First setup of leaderboard

Makefile

@@ -0,0 +1,13 @@
+.PHONY: style format
+
+
+style:
+	python -m black --line-length 119 .
+	python -m isort .
+	ruff check --fix .
+
+
+quality:
+	python -m black --check --line-length 119 .
+	python -m isort --check-only .
+	ruff check .

README.md

@@ -1,11 +1,21 @@
 ---
-title: Leaderboard
-emoji: 🐢
+title: LLM Leaderboard
+emoji: 🏆
 colorFrom: green
 colorTo: indigo
-sdk: static
+sdk: gradio
+sdk_version: 4.14.0
+app_file: app.py
 pinned: false
 license: apache-2.0
+fullWidth: true
+startup_duration_timeout: 1h
+space_ci:
+  private: true
+  secrets:
+  - HF_TOKEN
+  - H4_TOKEN
+tags:
+- leaderboard
+short_description: Evaluate open LLMs using arena-hard
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,188 @@
+import json
+import logging
+import os
+import subprocess
+import time
+
+import gradio as gr
+import pandas as pd
+from apscheduler.schedulers.background import BackgroundScheduler
+from gradio_leaderboard import Leaderboard, SelectColumns
+from gradio_space_ci import enable_space_ci
+from huggingface_hub import snapshot_download
+
+from src.display.about import (
+    FAQ_TEXT,
+    INTRODUCTION_TEXT,
+    LLM_BENCHMARKS_TEXT,
+    TITLE,
+)
+from src.display.css_html_js import custom_css
+from src.display.utils import (
+    # BENCHMARK_COLS,
+    AutoEvalColumn,
+    fields,
+)
+from src.envs import (
+    API,
+    EVAL_RESULTS_PATH,
+    H4_TOKEN,
+    REPO_ID,
+    RESET_JUDGEMENT_ENV,
+)
+
+os.environ['GRADIO_ANALYTICS_ENABLED']='false'
+
+# Configure logging
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+# Start ephemeral Spaces on PRs (see config in README.md)
+enable_space_ci()
+
+def restart_space():
+    API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)
+
+
+def time_diff_wrapper(func):
+    def wrapper(*args, **kwargs):
+        start_time = time.time()
+        result = func(*args, **kwargs)
+        end_time = time.time()
+        diff = end_time - start_time
+        logging.info(f"Time taken for {func.__name__}: {diff} seconds")
+        return result
+    return wrapper
+
+
+@time_diff_wrapper
+def download_dataset(repo_id, local_dir, repo_type="dataset", max_attempts=3, backoff_factor=1.5):
+    """Download dataset with exponential backoff retries."""
+    attempt = 0
+    while attempt < max_attempts:
+        try:
+            logging.info(f"Downloading {repo_id} to {local_dir}")
+            snapshot_download(
+                repo_id=repo_id,
+                local_dir=local_dir,
+                repo_type=repo_type,
+                tqdm_class=None,
+                etag_timeout=30,
+                max_workers=8,
+            )
+            logging.info("Download successful")
+            return
+        except Exception as e:
+            wait_time = backoff_factor ** attempt
+            logging.error(f"Error downloading {repo_id}: {e}, retrying in {wait_time}s")
+            time.sleep(wait_time)
+            attempt += 1
+    raise Exception(f"Failed to download {repo_id} after {max_attempts} attempts")
+
+def init_space(full_init: bool = True):
+    """Initializes the application space, loading only necessary data."""
+    if full_init:
+        # These downloads only occur on full initialization
+        # try:
+            # download_dataset(QUEUE_REPO, EVAL_REQUESTS_PATH)
+            # download_dataset(DYNAMIC_INFO_REPO, DYNAMIC_INFO_PATH)
+            download_dataset("Vikhrmodels/openbench-eval", EVAL_RESULTS_PATH)
+            # print(subprocess.Popen('ls src'))
+            subprocess.run(['rsync', '-avzP', '--ignore-existing', f'{EVAL_RESULTS_PATH[2:]}/external/*', 'src/gen/data/arena-hard-v0.1/model_answer/'])
+            subprocess.run(['rsync', '-avzP', '--ignore-existing', f'{EVAL_RESULTS_PATH[2:]}/model_judgment/*', 'src/gen/data/arena-hard-v0.1/model_judgement/'])
+        # except Exception:
+        #     restart_space()
+
+    # Always retrieve the leaderboard DataFrame
+    original_df = pd.DataFrame.from_records(json.load(open('eval-results/evals/upd.json','r')))
+
+
+    leaderboard_df = original_df.copy()
+
+
+    return leaderboard_df
+
+# Convert the environment variable "LEADERBOARD_FULL_INIT" to a boolean value, defaulting to True if the variable is not set.
+# This controls whether a full initialization should be performed.
+do_full_init = os.getenv("LEADERBOARD_FULL_INIT", "True") == "True"
+
+# Calls the init_space function with the `full_init` parameter determined by the `do_full_init` variable.
+# This initializes various DataFrames used throughout the application, with the level of initialization detail controlled by the `do_full_init` flag.
+leaderboard_df = init_space(full_init=do_full_init)
+
+demo = gr.Blocks(css=custom_css)
+with demo:
+    gr.HTML(TITLE)
+    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
+
+    with gr.Tabs(elem_classes="tab-buttons") as tabs:
+        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
+            pass
+            leaderboard = Leaderboard(
+                value=leaderboard_df,
+                datatype=[c.type for c in fields(AutoEvalColumn)],
+                select_columns=SelectColumns(
+                    default_selection=[
+                        c.name
+                        for c in fields(AutoEvalColumn)
+                        if c.displayed_by_default
+                    ],
+                    cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden or c.dummy],
+                    label="Select Columns to Display:",
+                ),
+                search_columns=[
+                    AutoEvalColumn.model.name,
+                    # AutoEvalColumn.fullname.name,
+                    # AutoEvalColumn.license.name
+                ],
+            )
+
+
+        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=3):
+            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
+
+        with gr.TabItem("❗FAQ", elem_id="llm-benchmark-tab-table", id=4):
+            gr.Markdown(FAQ_TEXT, elem_classes="markdown-text")
+
+        with gr.TabItem("🚀 Submit ", elem_id="llm-benchmark-tab-table", id=5):
+
+            with gr.Row():
+                gr.Markdown("# ✨ Submit your model here!", elem_classes="markdown-text")
+
+            with gr.Column():
+                    model_name_textbox = gr.Textbox(label="Model name")
+                    def upload_file(file):
+                        print(file.name)
+                        file_path = file.name.split('/')[-1] if '/' in file.name else file.name
+                        print(file_path)
+                        API.upload_file(path_or_fileobj=file.name,path_in_repo='./external/'+file_path,repo_id='Vikhrmodels/openbench-eval',repo_type='dataset')
+                        os.environ[RESET_JUDGEMENT_ENV] = '1'
+
+                        return file.name
+                    if model_name_textbox:
+                        file_output = gr.File()
+                        upload_button = gr.UploadButton("Click to Upload & Submit Answers", file_types=['*'], file_count="single")
+                        upload_button.upload(upload_file, upload_button, file_output)
+             
+# print(os.system('cd src/gen && ../../.venv/bin/python gen_judgment.py'))
+# print(os.system('cd src/gen/ && python show_result.py --output'))
+           
+def update_board():
+    need_reset = os.environ.get(RESET_JUDGEMENT_ENV)
+    if need_reset != '1':
+        return
+    
+    os.environ[RESET_JUDGEMENT_ENV] = '0'
+    
+    subprocess.run(['python','../gen/gen_judgement.py'])
+
+    subprocess.Popen('python3 ../gen/show_result.py --output')
+
+
+if __name__ == "__main__":
+    os.environ[RESET_JUDGEMENT_ENV] = '1'
+    
+    scheduler = BackgroundScheduler()
+    scheduler.add_job(update_board, "interval", minutes=10) 
+    scheduler.start()
+
+    demo.queue(default_concurrency_limit=40).launch(debug=True)

pyproject.toml

@@ -0,0 +1,54 @@
+[tool.ruff]
+line-length = 120
+target-version = "py312"
+include = ["*.py", "*.pyi", "**/pyproject.toml", "*.ipynb"]
+ignore=["I","EM","FBT","TRY003","S101","D101","D102","D103","D104","D105","G004","D107","FA102"]
+fixable=["ALL"]
+select=["ALL"]
+
+[tool.ruff.lint]
+select = ["E", "F"]
+fixable = ["ALL"]
+ignore = ["E501"] # line too long (black is taking care of this)
+
+[tool.isort]
+profile = "black"
+line_length = 119
+
+[tool.black]
+line-length = 119
+
+[tool.poetry]
+package-mode = false
+name = "open-llm-leaderboard"
+version = "0.1.0"
+description = ""
+authors = []
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = "3.12.1"
+apscheduler = "3.10.1"
+black = "23.11.0"
+click = "8.1.3"
+datasets = "2.14.5"
+huggingface-hub = ">=0.18.0"
+matplotlib = "3.8.4"
+numpy = "1.26.0"
+pandas = "2.2.2"
+plotly = "5.14.1"
+python-dateutil = "2.8.2"
+requests = "2.28.2"
+sentencepiece = "^0.2.0"
+tqdm = "4.65.0"
+transformers = "4.40.0"
+tokenizers = ">=0.15.0"
+gradio-space-ci = {git = "https://huggingface.co/spaces/Wauplin/gradio-space-ci", rev = "0.2.3"}
+gradio = " 4.20.0"
+isort = "^5.13.2"
+ruff = "^0.3.5"
+gradio-leaderboard = "0.0.8"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

requirements.txt

@@ -0,0 +1,18 @@
+APScheduler==3.10.1
+black==23.11.0
+click==8.1.3
+datasets==2.14.5
+huggingface-hub>=0.18.0
+matplotlib==3.8.4
+numpy==1.26.0
+pandas==2.2.2
+plotly==5.14.1
+python-dateutil==2.8.2
+requests==2.28.2
+sentencepiece
+tqdm==4.65.0
+transformers==4.40.0
+tokenizers>=0.15.0
+gradio-space-ci @ git+https://huggingface.co/spaces/Wauplin/gradio-space-ci@0.2.3 # CI !!!
+gradio==4.20.0
+gradio_leaderboard==0.0.8

src/display/about.py

@@ -0,0 +1,307 @@
+from src.display.utils import ModelType
+
+TITLE = """<h1 style="text-align:left;float:left; id="space-title">🤗 Open LLM Leaderboard</h1> <h3 style="text-align:left;float:left;> Track, rank and evaluate open LLMs and chatbots </h3>"""
+
+INTRODUCTION_TEXT = """
+"""
+
+icons = f"""
+- {ModelType.PT.to_str(" : ")} model: new, base models, trained on a given text corpora using masked modelling
+- {ModelType.CPT.to_str(" : ")} model: new, base models, continuously trained on further corpus (which may include IFT/chat data) using masked modelling
+- {ModelType.FT.to_str(" : ")} model: pretrained models finetuned on more data
+- {ModelType.chat.to_str(" : ")} model: chat like fine-tunes, either using IFT (datasets of task instruction), RLHF or DPO (changing the model loss a bit with an added policy), etc
+- {ModelType.merges.to_str(" : ")} model: merges or MoErges, models which have been merged or fused without additional fine-tuning. 
+"""
+LLM_BENCHMARKS_TEXT = """
+## ABOUT
+With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art.
+
+🤗 Submit a model for automated evaluation on the 🤗 GPU cluster on the "Submit" page!
+The leaderboard's backend runs the great [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) - read more details below!
+
+### Tasks 
+📈 We evaluate models on 6 key benchmarks using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks.
+
+- <a href="https://arxiv.org/abs/1803.05457" target="_blank">  AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions.
+- <a href="https://arxiv.org/abs/1905.07830" target="_blank">  HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
+- <a href="https://arxiv.org/abs/2009.03300" target="_blank">  MMLU </a>  (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
+- <a href="https://arxiv.org/abs/2109.07958" target="_blank">  TruthfulQA </a> (0-shot) - a test to measure a model's propensity to reproduce falsehoods commonly found online. Note: TruthfulQA is technically a 6-shot task in the Harness because each example is prepended with 6 Q/A pairs, even in the 0-shot setting.
+- <a href="https://arxiv.org/abs/1907.10641" target="_blank">  Winogrande </a> (5-shot) - an adversarial and difficult Winograd benchmark at scale, for commonsense reasoning.
+- <a href="https://arxiv.org/abs/2110.14168" target="_blank">  GSM8k </a> (5-shot) - diverse grade school math word problems to measure a model's ability to solve multi-step mathematical reasoning problems.
+
+For all these evaluations, a higher score is a better score.
+We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings.
+
+### Results
+You can find:
+- detailed numerical results in the `results` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/results
+- details on the input/outputs for the models in the `details` of each model, which you can access by clicking the 📄 emoji after the model name
+- community queries and running status in the `requests` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/requests
+
+If a model's name contains "Flagged", this indicates it has been flagged by the community, and should probably be ignored! Clicking the link will redirect you to the discussion about the model.
+
+---------------------------
+
+## REPRODUCIBILITY
+To reproduce our results, here are the commands you can run, using [this version](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463) of the Eleuther AI Harness:
+`python main.py --model=hf-causal-experimental --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>"`
+` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=1 --output_path=<output_path>`
+
+```
+python main.py --model=hf-causal-experimental \
+    --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>" \
+    --tasks=<task_list> \
+    --num_fewshot=<n_few_shot> \
+    --batch_size=1 \
+    --output_path=<output_path>
+```
+
+**Note:** We evaluate all models on a single node of 8 H100s, so the global batch size is 8 for each evaluation. If you don't use parallelism, adapt your batch size to fit.
+*You can expect results to vary slightly for different batch sizes because of padding.*
+
+The tasks and few shots parameters are:
+- ARC: 25-shot, *arc-challenge* (`acc_norm`)
+- HellaSwag: 10-shot, *hellaswag* (`acc_norm`)
+- TruthfulQA: 0-shot, *truthfulqa-mc* (`mc2`)
+- MMLU: 5-shot, *hendrycksTest-abstract_algebra,hendrycksTest-anatomy,hendrycksTest-astronomy,hendrycksTest-business_ethics,hendrycksTest-clinical_knowledge,hendrycksTest-college_biology,hendrycksTest-college_chemistry,hendrycksTest-college_computer_science,hendrycksTest-college_mathematics,hendrycksTest-college_medicine,hendrycksTest-college_physics,hendrycksTest-computer_security,hendrycksTest-conceptual_physics,hendrycksTest-econometrics,hendrycksTest-electrical_engineering,hendrycksTest-elementary_mathematics,hendrycksTest-formal_logic,hendrycksTest-global_facts,hendrycksTest-high_school_biology,hendrycksTest-high_school_chemistry,hendrycksTest-high_school_computer_science,hendrycksTest-high_school_european_history,hendrycksTest-high_school_geography,hendrycksTest-high_school_government_and_politics,hendrycksTest-high_school_macroeconomics,hendrycksTest-high_school_mathematics,hendrycksTest-high_school_microeconomics,hendrycksTest-high_school_physics,hendrycksTest-high_school_psychology,hendrycksTest-high_school_statistics,hendrycksTest-high_school_us_history,hendrycksTest-high_school_world_history,hendrycksTest-human_aging,hendrycksTest-human_sexuality,hendrycksTest-international_law,hendrycksTest-jurisprudence,hendrycksTest-logical_fallacies,hendrycksTest-machine_learning,hendrycksTest-management,hendrycksTest-marketing,hendrycksTest-medical_genetics,hendrycksTest-miscellaneous,hendrycksTest-moral_disputes,hendrycksTest-moral_scenarios,hendrycksTest-nutrition,hendrycksTest-philosophy,hendrycksTest-prehistory,hendrycksTest-professional_accounting,hendrycksTest-professional_law,hendrycksTest-professional_medicine,hendrycksTest-professional_psychology,hendrycksTest-public_relations,hendrycksTest-security_studies,hendrycksTest-sociology,hendrycksTest-us_foreign_policy,hendrycksTest-virology,hendrycksTest-world_religions* (average of all the results `acc`)
+- Winogrande: 5-shot, *winogrande* (`acc`)
+- GSM8k: 5-shot, *gsm8k* (`acc`)
+
+Side note on the baseline scores: 
+- for log-likelihood evaluation, we select the random baseline
+- for GSM8K, we select the score obtained in the paper after finetuning a 6B model on the full GSM8K training set for 50 epochs
+
+---------------------------
+
+## RESOURCES
+
+### Quantization
+To get more information about quantization, see:
+- 8 bits: [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration), [paper](https://arxiv.org/abs/2208.07339)
+- 4 bits: [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes), [paper](https://arxiv.org/abs/2305.14314)
+
+### Useful links
+- [Community resources](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/174)
+- [Collection of best models](https://huggingface.co/collections/open-llm-leaderboard/llm-leaderboard-best-models-652d6c7965a4619fb5c27a03)
+
+### Other cool leaderboards:
+- [LLM safety](https://huggingface.co/spaces/AI-Secure/llm-trustworthy-leaderboard)
+- [LLM performance](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)
+
+
+"""
+
+FAQ_TEXT = """
+
+## SUBMISSIONS
+My model requires `trust_remote_code=True`, can I submit it?
+- *We only support models that have been integrated into a stable version of the `transformers` library for automatic submission, as we don't want to run possibly unsafe code on our cluster.*
+
+What about models of type X? 
+- *We only support models that have been integrated into a stable version of the `transformers` library for automatic submission.*
+
+How can I follow when my model is launched?
+- *You can look for its request file [here](https://huggingface.co/datasets/open-llm-leaderboard/requests) and follow the status evolution, or directly in the queues above the submit form.*
+
+My model disappeared from all the queues, what happened?
+- *A model disappearing from all the queues usually means that there has been a failure. You can check if that is the case by looking for your model [here](https://huggingface.co/datasets/open-llm-leaderboard/requests).*
+
+What causes an evaluation failure?
+- *Most of the failures we get come from problems in the submissions (corrupted files, config problems, wrong parameters selected for eval ...), so we'll be grateful if you first make sure you have followed the steps in `About`. However, from time to time, we have failures on our side (hardware/node failures, problems with an update of our backend, connectivity problems ending up in the results not being saved, ...).*
+
+How can I report an evaluation failure?
+- *As we store the logs for all models, feel free to create an issue, **where you link to the requests file of your model** (look for it [here](https://huggingface.co/datasets/open-llm-leaderboard/requests/tree/main)), so we can investigate! If the model failed due to a problem on our side, we'll relaunch it right away!* 
+*Note: Please do not re-upload your model under a different name, it will not help*
+
+---------------------------
+
+## RESULTS
+What kind of information can I find?
+- *Let's imagine you are interested in the Yi-34B results. You have access to 3 different information categories:*
+      - *The [request file](https://huggingface.co/datasets/open-llm-leaderboard/requests/blob/main/01-ai/Yi-34B_eval_request_False_bfloat16_Original.json): it gives you information about the status of the evaluation*
+      - *The [aggregated results folder](https://huggingface.co/datasets/open-llm-leaderboard/results/tree/main/01-ai/Yi-34B): it gives you aggregated scores, per experimental run*
+      - *The [details dataset](https://huggingface.co/datasets/open-llm-leaderboard/details_01-ai__Yi-34B/tree/main): it gives you the full details (scores and examples for each task and a given model)*
+
+
+Why do models appear several times in the leaderboard? 
+- *We run evaluations with user-selected precision and model commit. Sometimes, users submit specific models at different commits and at different precisions (for example, in float16 and 4bit to see how quantization affects performance). You should be able to verify this by displaying the `precision` and `model sha` columns in the display. If, however, you see models appearing several times with the same precision and hash commit, this is not normal.*
+
+What is this concept of "flagging"?
+- *This mechanism allows users to report models that have unfair performance on the leaderboard. This contains several categories: exceedingly good results on the leaderboard because the model was (maybe accidentally) trained on the evaluation data, models that are copies of other models not attributed properly, etc.*
+
+My model has been flagged improperly, what can I do?
+- *Every flagged model has a discussion associated with it - feel free to plead your case there, and we'll see what to do together with the community.*
+
+---------------------------
+
+## HOW TO SEARCH FOR A MODEL
+Search for models in the leaderboard by:
+1. Name, e.g., *model_name*
+2. Multiple names, separated by `;`, e.g., *model_name1;model_name2*
+3. License, prefix with `license:`, e.g., *license: MIT*
+4. Combination of name and license, order is irrelevant, e.g., *model_name; license: cc-by-sa-4.0*
+
+---------------------------
+
+## EDITING SUBMISSIONS
+I upgraded my model and want to re-submit, how can I do that?
+- *Please open an issue with the precise name of your model, and we'll remove your model from the leaderboard so you can resubmit. You can also resubmit directly with the new commit hash!* 
+
+I need to rename my model, how can I do that?
+- *You can use @Weyaxi 's [super cool tool](https://huggingface.co/spaces/Weyaxi/open-llm-leaderboard-renamer) to request model name changes, then open a discussion where you link to the created pull request, and we'll check them and merge them as needed.*
+
+---------------------------
+
+## OTHER
+Why do you differentiate between pretrained, continuously pretrained, fine-tuned, merges, etc?
+- *These different models do not play in the same categories, and therefore need to be separated for fair comparison. Base pretrained models are the most interesting for the community, as they are usually good models to fine-tune later on - any jump in performance from a pretrained model represents a true improvement on the SOTA. 
+Fine-tuned and IFT/RLHF/chat models usually have better performance, but the latter might be more sensitive to system prompts, which we do not cover at the moment in the Open LLM Leaderboard. 
+Merges and moerges have artificially inflated performance on test sets, which is not always explainable, and does not always apply to real-world situations.*
+
+What should I use the leaderboard for?
+- *We recommend using the leaderboard for 3 use cases: 1) getting an idea of the state of open pretrained models, by looking only at the ranks and score of this category; 2) experimenting with different fine-tuning methods, datasets, quantization techniques, etc, and comparing their score in a reproducible setup, and 3) checking the performance of a model of interest to you, wrt to other models of its category.*
+
+Why don't you display closed-source model scores? 
+- *This is a leaderboard for Open models, both for philosophical reasons (openness is cool) and for practical reasons: we want to ensure that the results we display are accurate and reproducible, but 1) commercial closed models can change their API thus rendering any scoring at a given time incorrect 2) we re-run everything on our cluster to ensure all models are run on the same setup and you can't do that for these models.*
+
+I have an issue with accessing the leaderboard through the Gradio API
+- *Since this is not the recommended way to access the leaderboard, we won't provide support for this, but you can look at tools provided by the community for inspiration!*
+
+I have another problem, help!
+- *Please open an issue in the discussion tab, and we'll do our best to help you in a timely manner :) *
+"""
+
+
+EVALUATION_QUEUE_TEXT = f"""
+# Evaluation Queue for the 🤗 Open LLM Leaderboard
+
+Models added here will be automatically evaluated on the 🤗 cluster.
+
+## Don't forget to read the FAQ and the About tabs for more information!
+
+## First steps before submitting a model
+
+### 1) Make sure you can load your model and tokenizer using AutoClasses:
+```python
+from transformers import AutoConfig, AutoModel, AutoTokenizer
+config = AutoConfig.from_pretrained("your model name", revision=revision)
+model = AutoModel.from_pretrained("your model name", revision=revision)
+tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
+```
+If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded.
+
+Note: make sure your model is public!
+Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
+
+### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
+It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
+
+### 3) Make sure your model has an open license!
+This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
+
+### 4) Fill up your model card
+When we add extra information about models to the leaderboard, it will be automatically taken from the model card
+
+### 5) Select the correct precision
+Not all models are converted properly from `float16` to `bfloat16`, and selecting the wrong precision can sometimes cause evaluation error (as loading a `bf16` model in `fp16` can sometimes generate NaNs, depending on the weight range).
+
+<b>Note:</b> Please be advised that when submitting, git <b>branches</b> and <b>tags</b> will be strictly tied to the <b>specific commit</b> present at the time of submission. This ensures revision consistency.
+## Model types
+{icons}
+"""
+
+CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
+CITATION_BUTTON_TEXT = r"""
+@misc{open-llm-leaderboard,
+  author = {Edward Beeching and Clémentine Fourrier and Nathan Habib and Sheon Han and Nathan Lambert and Nazneen Rajani and Omar Sanseviero and Lewis Tunstall and Thomas Wolf},
+  title = {Open LLM Leaderboard},
+  year = {2023},
+  publisher = {Hugging Face},
+  howpublished = "\url{https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard}"
+}
+@software{eval-harness,
+  author       = {Gao, Leo and
+                  Tow, Jonathan and
+                  Biderman, Stella and
+                  Black, Sid and
+                  DiPofi, Anthony and
+                  Foster, Charles and
+                  Golding, Laurence and
+                  Hsu, Jeffrey and
+                  McDonell, Kyle and
+                  Muennighoff, Niklas and
+                  Phang, Jason and
+                  Reynolds, Laria and
+                  Tang, Eric and
+                  Thite, Anish and
+                  Wang, Ben and
+                  Wang, Kevin and
+                  Zou, Andy},
+  title        = {A framework for few-shot language model evaluation},
+  month        = sep,
+  year         = 2021,
+  publisher    = {Zenodo},
+  version      = {v0.0.1},
+  doi          = {10.5281/zenodo.5371628},
+  url          = {https://doi.org/10.5281/zenodo.5371628}
+}
+@misc{clark2018think,
+      title={Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge},
+      author={Peter Clark and Isaac Cowhey and Oren Etzioni and Tushar Khot and Ashish Sabharwal and Carissa Schoenick and Oyvind Tafjord},
+      year={2018},
+      eprint={1803.05457},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI}
+}
+@misc{zellers2019hellaswag,
+      title={HellaSwag: Can a Machine Really Finish Your Sentence?},
+      author={Rowan Zellers and Ari Holtzman and Yonatan Bisk and Ali Farhadi and Yejin Choi},
+      year={2019},
+      eprint={1905.07830},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+@misc{hendrycks2021measuring,
+      title={Measuring Massive Multitask Language Understanding},
+      author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt},
+      year={2021},
+      eprint={2009.03300},
+      archivePrefix={arXiv},
+      primaryClass={cs.CY}
+}
+@misc{lin2022truthfulqa,
+      title={TruthfulQA: Measuring How Models Mimic Human Falsehoods},
+      author={Stephanie Lin and Jacob Hilton and Owain Evans},
+      year={2022},
+      eprint={2109.07958},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+@misc{DBLP:journals/corr/abs-1907-10641,
+      title={{WINOGRANDE:} An Adversarial Winograd Schema Challenge at Scale},
+      author={Keisuke Sakaguchi and Ronan Le Bras and Chandra Bhagavatula and Yejin Choi},
+      year={2019},
+      eprint={1907.10641},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+@misc{DBLP:journals/corr/abs-2110-14168,
+      title={Training Verifiers to Solve Math Word Problems},
+      author={Karl Cobbe and
+                  Vineet Kosaraju and
+                  Mohammad Bavarian and
+                  Mark Chen and
+                  Heewoo Jun and
+                  Lukasz Kaiser and
+                  Matthias Plappert and
+                  Jerry Tworek and
+                  Jacob Hilton and
+                  Reiichiro Nakano and
+                  Christopher Hesse and
+                  John Schulman},
+      year={2021},
+      eprint={2110.14168},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+"""

src/display/css_html_js.py

@@ -0,0 +1,91 @@
+custom_css = """
+/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
+table td:first-child,
+table th:first-child {
+    max-width: 400px;
+    overflow: auto;
+    white-space: nowrap;
+}
+
+/* Full width space */
+.gradio-container {
+  max-width: 95%!important;
+}
+
+/* Text style and margins */
+.markdown-text {
+    font-size: 16px !important;
+}
+
+#models-to-add-text {
+    font-size: 18px !important;
+}
+
+#citation-button span {
+    font-size: 16px !important;
+}
+
+#citation-button textarea {
+    font-size: 16px !important;
+}
+
+#citation-button > label > button {
+    margin: 6px;
+    transform: scale(1.3);
+}
+
+#search-bar-table-box > div:first-child {
+    background: none;
+    border: none;
+}
+
+#search-bar {
+    padding: 0px;
+}
+
+.tab-buttons button {
+    font-size: 20px;
+}
+
+/* Filters style */
+#filter_type{
+    border: 0;
+    padding-left: 0;
+    padding-top: 0;
+}
+#filter_type label {
+    display: flex;
+}
+#filter_type label > span{
+    margin-top: var(--spacing-lg);
+    margin-right: 0.5em;
+}
+#filter_type label > .wrap{
+    width: 103px;
+}
+#filter_type label > .wrap .wrap-inner{  
+    padding: 2px;
+}
+#filter_type label > .wrap .wrap-inner input{
+    width: 1px
+}
+#filter-columns-type{
+    border:0;
+    padding:0.5;
+}
+#filter-columns-size{
+    border:0;
+    padding:0.5;
+}
+#box-filter > .form{
+    border: 0
+}
+"""
+
+get_window_url_params = """
+    function(url_params) {
+        const params = new URLSearchParams(window.location.search);
+        url_params = Object.fromEntries(params);
+        return url_params;
+    }
+    """

src/display/formatting.py

@@ -0,0 +1,36 @@
+from huggingface_hub import HfApi
+
+API = HfApi()
+
+
+def model_hyperlink(link, model_name):
+    return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
+
+
+def make_clickable_model(model_name):
+    link = f"https://huggingface.co/{model_name}"
+
+    details_model_name = model_name.replace("/", "__")
+    details_link = f"https://huggingface.co/datasets/open-llm-leaderboard/details_{details_model_name}"
+
+    return model_hyperlink(link, model_name) + "  " + model_hyperlink(details_link, "📑")
+
+
+def styled_error(error):
+    return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"
+
+
+def styled_warning(warn):
+    return f"<p style='color: orange; font-size: 20px; text-align: center;'>{warn}</p>"
+
+
+def styled_message(message):
+    return f"<p style='color: green; font-size: 20px; text-align: center;'>{message}</p>"
+
+
+def has_no_nan_values(df, columns):
+    return df[columns].notna().all(axis=1)
+
+
+def has_nan_values(df, columns):
+    return df[columns].isna().any(axis=1)

src/display/utils.py

@@ -0,0 +1,233 @@
+from dataclasses import dataclass, make_dataclass
+from enum import Enum
+import json
+import logging
+from datetime import datetime
+import pandas as pd
+
+
+# Configure logging
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+def parse_datetime(datetime_str):
+    formats = [
+        "%Y-%m-%dT%H-%M-%S.%f",  # Format with dashes
+        "%Y-%m-%dT%H:%M:%S.%f",  # Standard format with colons
+        "%Y-%m-%dT%H %M %S.%f",  # Spaces as separator
+    ]
+
+    for fmt in formats:
+        try:
+            return datetime.strptime(datetime_str, fmt)
+        except ValueError:
+            continue
+    # in rare cases set unix start time for files with incorrect time (legacy files)
+    logging.error(f"No valid date format found for: {datetime_str}")
+    return datetime(1970, 1, 1)
+
+def load_json_data(file_path):
+    """Safely load JSON data from a file."""
+    try:
+        with open(file_path, "r") as file:
+            return json.load(file)
+    except json.JSONDecodeError:
+        print(f"Error reading JSON from {file_path}")
+        return None  # Or raise an exception
+
+
+def fields(raw_class):
+    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
+
+
+@dataclass
+class Task:
+    benchmark: str
+    metric: str
+    col_name: str
+
+
+class Tasks(Enum):
+    arc = Task("arc:challenge", "acc_norm", "ARC")
+    hellaswag = Task("hellaswag", "acc_norm", "HellaSwag")
+    mmlu = Task("hendrycksTest", "acc", "MMLU")
+    truthfulqa = Task("truthfulqa:mc", "mc2", "TruthfulQA")
+    winogrande = Task("winogrande", "acc", "Winogrande")
+    gsm8k = Task("gsm8k", "acc", "GSM8K")
+
+
+# These classes are for user facing column names,
+# to avoid having to change them all around the code
+# when a modif is needed
+@dataclass(frozen=True)
+class ColumnContent:
+    name: str
+    type: str
+    displayed_by_default: bool
+    hidden: bool = False
+    never_hidden: bool = False
+    dummy: bool = False
+
+
+auto_eval_column_dict = []
+# Init
+# auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
+auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("model", "markdown", True, never_hidden=True)])
+# # Scores
+auto_eval_column_dict.append(["score", ColumnContent, ColumnContent("score", "number", True)])
+# for task in Tasks:
+#     auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
+# # Model information
+# auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
+# auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
+# auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
+# auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
+# auto_eval_column_dict.append(["merged", ColumnContent, ColumnContent("Merged", "bool", False)])
+# auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
+# auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
+# auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
+# auto_eval_column_dict.append(
+#     ["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)]
+# )
+# auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
+# auto_eval_column_dict.append(["not_flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
+# auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
+# Dummy column for the search bar (hidden by the custom CSS)
+# auto_eval_column_dict.append(["tokens", ColumnContent, ColumnContent("avg_tokens", "str", False, dummy=True)])
+
+# We use make dataclass to dynamically fill the scores from Tasks
+AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
+
+
+
+@dataclass(frozen=True)
+class EvalQueueColumn:  # Queue column
+    model = ColumnContent("model", "markdown", True)
+    # revision = ColumnContent("revision", "str", True)
+    # private = ColumnContent("private", "bool", True)
+    # precision = ColumnContent("precision", "str", True)
+    # weight_type = ColumnContent("weight_type", "str", "Original")
+    # status = ColumnContent("status", "str", True)
+
+
+baseline_row = {
+    AutoEvalColumn.model.name: "<p>Baseline</p>",
+    # AutoEvalColumn.revision.name: "N/A",
+    # AutoEvalColumn.precision.name: None,
+    # AutoEvalColumn.merged.name: False,
+    # AutoEvalColumn.average.name: 31.0,
+    # AutoEvalColumn.arc.name: 25.0,
+    # AutoEvalColumn.hellaswag.name: 25.0,
+    # AutoEvalColumn.mmlu.name: 25.0,
+    # AutoEvalColumn.truthfulqa.name: 25.0,
+    # AutoEvalColumn.winogrande.name: 50.0,
+    # AutoEvalColumn.gsm8k.name: 0.21,
+    # AutoEvalColumn.fullname.name: "baseline",
+    # AutoEvalColumn.model_type.name: "",
+    # AutoEvalColumn.not_flagged.name: False,
+}
+
+# Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
+# ARC human baseline is 0.80 (source: https://lab42.global/arc/)
+# HellaSwag human baseline is 0.95 (source: https://deepgram.com/learn/hellaswag-llm-benchmark-guide)
+# MMLU human baseline is 0.898 (source: https://openreview.net/forum?id=d7KBjmI3GmQ)
+# TruthfulQA human baseline is 0.94(source: https://arxiv.org/pdf/2109.07958.pdf)
+# Winogrande: https://leaderboard.allenai.org/winogrande/submissions/public
+# GSM8K: paper
+# Define the human baselines
+human_baseline_row = {
+    AutoEvalColumn.model.name: "<p>Human performance</p>",
+    # AutoEvalColumn.revision.name: "N/A",
+    # AutoEvalColumn.precision.name: None,
+    # AutoEvalColumn.average.name: 92.75,
+    # AutoEvalColumn.merged.name: False,
+    # AutoEvalColumn.arc.name: 80.0,
+    # AutoEvalColumn.hellaswag.name: 95.0,
+    # AutoEvalColumn.mmlu.name: 89.8,
+    # AutoEvalColumn.truthfulqa.name: 94.0,
+    # AutoEvalColumn.winogrande.name: 94.0,
+    # AutoEvalColumn.gsm8k.name: 100,
+    # AutoEvalColumn.fullname.name: "human_baseline",
+    # AutoEvalColumn.model_type.name: "",
+    # AutoEvalColumn.not_flagged.name: False,
+}
+
+
+@dataclass
+class ModelDetails:
+    name: str
+    symbol: str = ""  # emoji, only for the model type
+
+
+class ModelType(Enum):
+    PT = ModelDetails(name="pretrained", symbol="🟢")
+    CPT = ModelDetails(name="continuously pretrained", symbol="🟩")
+    FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶")
+    chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬")
+    merges = ModelDetails(name="base merges and moerges", symbol="🤝")
+    Unknown = ModelDetails(name="", symbol="?")
+
+    def to_str(self, separator=" "):
+        return f"{self.value.symbol}{separator}{self.value.name}"
+
+    @staticmethod
+    def from_str(type):
+        if "fine-tuned" in type or "🔶" in type:
+            return ModelType.FT
+        if "continously pretrained" in type or "🟩" in type:
+            return ModelType.CPT
+        if "pretrained" in type or "🟢" in type:
+            return ModelType.PT
+        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
+            return ModelType.chat
+        if "merge" in type or "🤝" in type:
+            return ModelType.merges
+        return ModelType.Unknown
+
+
+class WeightType(Enum):
+    Adapter = ModelDetails("Adapter")
+    Original = ModelDetails("Original")
+    Delta = ModelDetails("Delta")
+
+
+class Precision(Enum):
+    float16 = ModelDetails("float16")
+    bfloat16 = ModelDetails("bfloat16")
+    qt_8bit = ModelDetails("8bit")
+    qt_4bit = ModelDetails("4bit")
+    qt_GPTQ = ModelDetails("GPTQ")
+    Unknown = ModelDetails("?")
+
+    def from_str(precision):
+        if precision in ["torch.float16", "float16"]:
+            return Precision.float16
+        if precision in ["torch.bfloat16", "bfloat16"]:
+            return Precision.bfloat16
+        if precision in ["8bit"]:
+            return Precision.qt_8bit
+        if precision in ["4bit"]:
+            return Precision.qt_4bit
+        if precision in ["GPTQ", "None"]:
+            return Precision.qt_GPTQ
+        return Precision.Unknown
+
+
+# Column selection
+COLS = [c.name for c in fields(AutoEvalColumn)]
+TYPES = [c.type for c in fields(AutoEvalColumn)]
+
+EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
+EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
+
+# BENCHMARK_COLS = [t.value.col_name for t in Tasks]
+
+NUMERIC_INTERVALS = {
+    "?": pd.Interval(-1, 0, closed="right"),
+    "~1.5": pd.Interval(0, 2, closed="right"),
+    "~3": pd.Interval(2, 4, closed="right"),
+    "~7": pd.Interval(4, 9, closed="right"),
+    "~13": pd.Interval(9, 20, closed="right"),
+    "~35": pd.Interval(20, 45, closed="right"),
+    "~60": pd.Interval(45, 70, closed="right"),
+    "70+": pd.Interval(70, 10000, closed="right"),
+}

src/envs.py

@@ -0,0 +1,47 @@
+import os
+
+from huggingface_hub import HfApi
+
+# clone / pull the lmeh eval data
+H4_TOKEN = os.environ.get("H4_TOKEN", None)
+
+REPO_ID = "HuggingFaceH4/open_llm_leaderboard"
+QUEUE_REPO = "open-llm-leaderboard/requests"
+DYNAMIC_INFO_REPO = "open-llm-leaderboard/dynamic_model_information"
+RESULTS_REPO = "open-llm-leaderboard/results"
+
+PRIVATE_QUEUE_REPO = "open-llm-leaderboard/private-requests"
+PRIVATE_RESULTS_REPO = "open-llm-leaderboard/private-results"
+
+IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
+
+HF_HOME = os.getenv("HF_HOME", ".")
+
+# Check HF_HOME write access
+print(f"Initial HF_HOME set to: {HF_HOME}")
+
+if not os.access(HF_HOME, os.W_OK):
+    print(f"No write access to HF_HOME: {HF_HOME}. Resetting to current directory.")
+    HF_HOME = "."
+    os.environ["HF_HOME"] = HF_HOME
+else:
+    print("Write access confirmed for HF_HOME")
+
+EVAL_REQUESTS_PATH = os.path.join(HF_HOME, "eval-queue")
+EVAL_RESULTS_PATH = os.path.join(HF_HOME, "eval-results")
+DYNAMIC_INFO_PATH = os.path.join(HF_HOME, "dynamic-info")
+DYNAMIC_INFO_FILE_PATH = os.path.join(DYNAMIC_INFO_PATH, "model_infos.json")
+
+EVAL_REQUESTS_PATH_PRIVATE = "eval-queue-private"
+EVAL_RESULTS_PATH_PRIVATE = "eval-results-private"
+
+PATH_TO_COLLECTION = "open-llm-leaderboard/llm-leaderboard-best-models-652d6c7965a4619fb5c27a03"
+
+# Rate limit variables
+RATE_LIMIT_PERIOD = 7
+RATE_LIMIT_QUOTA = 5
+HAS_HIGHER_RATE_LIMIT = ["TheBloke"]
+
+RESET_JUDGEMENT_ENV = "RESET_JUDGEMENT"
+
+API = HfApi(token=H4_TOKEN)

src/gen/arena_hard_leaderboard_20240514.json

@@ -0,0 +1,329 @@
+[
+    {
+        "results":[
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0
+        ],
+        "model":"gpt-3.5-turbo-0125",
+        "score":50.0,
+        "lower":50.0,
+        "upper":50.0,
+        "avg_tokens":0.0
+    },
+    {
+        "results":[
+            855.5644665503,
+            859.0709454157,
+            865.0434024226,
+            860.399655762,
+            855.1731508697,
+            855.5326400531,
+            866.7819454641,
+            858.5219875589,
+            861.4603125434,
+            859.8350548067,
+            862.7609222876,
+            854.2414273092,
+            862.374147169,
+            863.1792770928,
+            865.2996605704,
+            864.8988771163,
+            867.0356240274,
+            871.6157440982,
+            861.9225322393,
+            864.7557130348,
+            853.284444198,
+            851.7087385877,
+            871.482425846,
+            866.6122634027,
+            852.7157509126,
+            859.7938560994,
+            874.1682886992,
+            855.4589887037,
+            850.0205093168,
+            875.7282859976,
+            865.3647024942,
+            856.1797064852,
+            867.6238850835,
+            857.7097671655,
+            874.4978660071,
+            857.5650653089,
+            890.8852955482,
+            855.6426165155,
+            859.3456423505,
+            857.4854945486,
+            880.1901418236,
+            849.6103242372,
+            871.0458800663,
+            877.4244267245,
+            875.3479511716,
+            859.1269918194,
+            857.8015195801,
+            868.2750694028,
+            868.0957706924,
+            870.6012679715,
+            862.269673472,
+            864.2488571071,
+            874.1624601722,
+            863.1194231025,
+            857.1192986285,
+            862.0030926827,
+            861.5474187298,
+            880.5566205251,
+            861.7223684538,
+            874.9512628918,
+            858.7260910186,
+            871.4133525673,
+            866.2715335516,
+            861.3256361213,
+            866.9022358038,
+            867.5601382523,
+            864.5272121008,
+            866.7782194777,
+            865.4086246736,
+            870.0314924292,
+            855.3587976891,
+            851.5511568095,
+            863.2094645624,
+            861.0624318318,
+            848.5397354473,
+            857.9432204946,
+            861.2370229881,
+            878.2964116149,
+            857.9909782749,
+            871.9069179589,
+            860.2445059252,
+            850.4012745111,
+            866.7922558028,
+            862.2175409513,
+            856.8494155845,
+            856.4641060792,
+            878.905415424,
+            851.8853822745,
+            859.2360763272,
+            869.1579952553,
+            855.2369472583,
+            859.2009612357,
+            876.2027799847,
+            849.6362696273,
+            865.1318475963,
+            855.8791178271,
+            873.3916447336,
+            867.1797828548,
+            865.1613697328,
+            875.1689869302
+        ],
+        "model":"gigachat_pro",
+        "score":31.37,
+        "lower":29.64,
+        "upper":33.33,
+        "avg_tokens":0.0
+    },
+    {
+        "results":[
+            726.6208252619,
+            738.5741612323,
+            734.1011761886,
+            729.5571514643,
+            728.758372467,
+            733.7900136425,
+            719.043685497,
+            714.8370789545,
+            725.8752720444,
+            715.266084892,
+            727.2017077065,
+            739.3798608124,
+            719.6304899658,
+            734.0546251412,
+            718.4924449088,
+            721.0729415472,
+            738.5699274129,
+            723.7105361329,
+            728.2971721354,
+            737.8461934603,
+            748.9971545908,
+            713.1462726999,
+            720.2960317186,
+            727.2517234335,
+            694.2654473149,
+            735.6639839406,
+            730.5016731736,
+            734.4551919945,
+            728.8931636911,
+            717.6726330463,
+            733.3721052861,
+            725.7981758416,
+            731.0409312559,
+            715.3647090465,
+            737.7875979517,
+            729.3512200797,
+            715.9010959711,
+            722.2116159282,
+            724.6752254921,
+            718.5749125859,
+            723.0132896162,
+            732.3587564613,
+            740.6268654101,
+            724.6297632896,
+            743.701641735,
+            723.5736702859,
+            731.9752231934,
+            722.3929635211,
+            721.9705147906,
+            738.9123529498,
+            733.7609432817,
+            724.1850017217,
+            727.8550112565,
+            731.3315308989,
+            722.5721295254,
+            729.8940208849,
+            735.9873637973,
+            730.6501947523,
+            702.8268457509,
+            732.6491227137,
+            736.225411771,
+            745.6156113918,
+            721.0912474577,
+            736.2254117629,
+            732.9674153867,
+            723.0966793643,
+            718.0704518208,
+            722.2852812675,
+            745.1185090985,
+            736.9690722951,
+            742.6306627437,
+            733.1555506911,
+            721.7491525609,
+            723.0795022704,
+            717.9478748234,
+            726.703609728,
+            725.3073844986,
+            722.2116156669,
+            720.1865370325,
+            731.5240457448,
+            737.0781670626,
+            708.356058121,
+            730.3511179714,
+            727.5035049316,
+            706.4191731996,
+            734.2333848904,
+            736.5196621633,
+            724.9647865416,
+            718.7060814362,
+            722.5615781913,
+            731.6666527735,
+            722.1914533305,
+            719.1795542579,
+            730.3223324585,
+            724.1322488355,
+            734.6332090556,
+            716.1292305518,
+            726.7846008592,
+            717.027778133,
+            728.6562483681
+        ],
+        "model":"gigachat_lite",
+        "score":17.2,
+        "lower":15.65,
+        "upper":18.68,
+        "avg_tokens":276.0
+    }
+]

src/gen/arena_hard_leaderboard_20240515.json

@@ -0,0 +1,329 @@
+[
+    {
+        "results":[
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0,
+            1000.0
+        ],
+        "model":"gpt-3.5-turbo-0125",
+        "score":50.0,
+        "lower":50.0,
+        "upper":50.0,
+        "avg_tokens":0.0
+    },
+    {
+        "results":[
+            855.5644665503,
+            859.0709454157,
+            865.0434024226,
+            860.399655762,
+            855.1731508697,
+            855.5326400531,
+            866.7819454641,
+            858.5219875589,
+            861.4603125434,
+            859.8350548067,
+            862.7609222876,
+            854.2414273092,
+            862.374147169,
+            863.1792770928,
+            865.2996605704,
+            864.8988771163,
+            867.0356240274,
+            871.6157440982,
+            861.9225322393,
+            864.7557130348,
+            853.284444198,
+            851.7087385877,
+            871.482425846,
+            866.6122634027,
+            852.7157509126,
+            859.7938560994,
+            874.1682886992,
+            855.4589887037,
+            850.0205093168,
+            875.7282859976,
+            865.3647024942,
+            856.1797064852,
+            867.6238850835,
+            857.7097671655,
+            874.4978660071,
+            857.5650653089,
+            890.8852955482,
+            855.6426165155,
+            859.3456423505,
+            857.4854945486,
+            880.1901418236,
+            849.6103242372,
+            871.0458800663,
+            877.4244267245,
+            875.3479511716,
+            859.1269918194,
+            857.8015195801,
+            868.2750694028,
+            868.0957706924,
+            870.6012679715,
+            862.269673472,
+            864.2488571071,
+            874.1624601722,
+            863.1194231025,
+            857.1192986285,
+            862.0030926827,
+            861.5474187298,
+            880.5566205251,
+            861.7223684538,
+            874.9512628918,
+            858.7260910186,
+            871.4133525673,
+            866.2715335516,
+            861.3256361213,
+            866.9022358038,
+            867.5601382523,
+            864.5272121008,
+            866.7782194777,
+            865.4086246736,
+            870.0314924292,
+            855.3587976891,
+            851.5511568095,
+            863.2094645624,
+            861.0624318318,
+            848.5397354473,
+            857.9432204946,
+            861.2370229881,
+            878.2964116149,
+            857.9909782749,
+            871.9069179589,
+            860.2445059252,
+            850.4012745111,
+            866.7922558028,
+            862.2175409513,
+            856.8494155845,
+            856.4641060792,
+            878.905415424,
+            851.8853822745,
+            859.2360763272,
+            869.1579952553,
+            855.2369472583,
+            859.2009612357,
+            876.2027799847,
+            849.6362696273,
+            865.1318475963,
+            855.8791178271,
+            873.3916447336,
+            867.1797828548,
+            865.1613697328,
+            875.1689869302
+        ],
+        "model":"gigachat_pro",
+        "score":31.37,
+        "lower":29.64,
+        "upper":33.33,
+        "avg_tokens":0.0
+    },
+    {
+        "results":[
+            726.6208252619,
+            738.5741612323,
+            734.1011761886,
+            729.5571514643,
+            728.758372467,
+            733.7900136425,
+            719.043685497,
+            714.8370789545,
+            725.8752720444,
+            715.266084892,
+            727.2017077065,
+            739.3798608124,
+            719.6304899658,
+            734.0546251412,
+            718.4924449088,
+            721.0729415472,
+            738.5699274129,
+            723.7105361329,
+            728.2971721354,
+            737.8461934603,
+            748.9971545908,
+            713.1462726999,
+            720.2960317186,
+            727.2517234335,
+            694.2654473149,
+            735.6639839406,
+            730.5016731736,
+            734.4551919945,
+            728.8931636911,
+            717.6726330463,
+            733.3721052861,
+            725.7981758416,
+            731.0409312559,
+            715.3647090465,
+            737.7875979517,
+            729.3512200797,
+            715.9010959711,
+            722.2116159282,
+            724.6752254921,
+            718.5749125859,
+            723.0132896162,
+            732.3587564613,
+            740.6268654101,
+            724.6297632896,
+            743.701641735,
+            723.5736702859,
+            731.9752231934,
+            722.3929635211,
+            721.9705147906,
+            738.9123529498,
+            733.7609432817,
+            724.1850017217,
+            727.8550112565,
+            731.3315308989,
+            722.5721295254,
+            729.8940208849,
+            735.9873637973,
+            730.6501947523,
+            702.8268457509,
+            732.6491227137,
+            736.225411771,
+            745.6156113918,
+            721.0912474577,
+            736.2254117629,
+            732.9674153867,
+            723.0966793643,
+            718.0704518208,
+            722.2852812675,
+            745.1185090985,
+            736.9690722951,
+            742.6306627437,
+            733.1555506911,
+            721.7491525609,
+            723.0795022704,
+            717.9478748234,
+            726.703609728,
+            725.3073844986,
+            722.2116156669,
+            720.1865370325,
+            731.5240457448,
+            737.0781670626,
+            708.356058121,
+            730.3511179714,
+            727.5035049316,
+            706.4191731996,
+            734.2333848904,
+            736.5196621633,
+            724.9647865416,
+            718.7060814362,
+            722.5615781913,
+            731.6666527735,
+            722.1914533305,
+            719.1795542579,
+            730.3223324585,
+            724.1322488355,
+            734.6332090556,
+            716.1292305518,
+            726.7846008592,
+            717.027778133,
+            728.6562483681
+        ],
+        "model":"gigachat_lite",
+        "score":17.2,
+        "lower":15.65,
+        "upper":18.68,
+        "avg_tokens":276.0
+    }
+]

src/gen/config/api_config.yaml

@@ -0,0 +1,203 @@
+# name: str
+#     model_name: str
+#     endpoints: default to null
+#         - api_base: str
+#           api_key: str optional (required if no api_key_ENV)
+#           api_key_ENV: str optional (ENV name to store the token secret)
+#           api_version: str optional (only for azure)
+#     api_type: str
+#     tokenizer: str optional (to optimize token limits)
+#     parallel: int
+
+gpt-4-1106-preview:
+    model_name: gpt-4-1106-preview
+    endpoints:
+        - api_base: https://cgiaura-openai-trainning.openai.azure.com
+          api_key_ENV: GPT_4_TOKEN
+          api_version: 2024-02-15-preview
+    api_type: azure
+    parallel: 5
+
+gpt-3.5-turbo-0125:
+    model_name: gpt-3.5-turbo-0125
+    endpoints:
+        - api_base: https://api.openai.com/v1/
+          api_key_ENV: GPT_3_TOKEN
+    api_type: openai
+    parallel: 6
+
+gpt-3.5-turbo-0125-ru-sys:
+    model_name: gpt-3.5-turbo-0125
+    endpoints:
+        - api_base: https://api.openai.com/v1/
+          api_key_ENV: GPT_3_TOKEN
+    system_prompt: You are a helpful assistant. Answer on Russian.
+    api_type: openai
+    parallel: 6
+
+yandex_gpt_pro:
+    model_name: yandexgpt
+    endpoints:
+        - catalog_id: b1gk1i41eeb97a5s68c7
+          iam_token_ENV: YANDEX_GPT_TOKEN
+    api_type: yandex
+    parallel: 2
+
+gigachat_lite:
+    model_name: GigaChat
+    endpoints:
+          auth_token_ENV: GIGACHAT_GPT_TOKEN
+    api_type: gigachat
+    parallel: 1
+
+gigachat_pro:
+    model_name: GigaChat-Pro
+    endpoints:
+          auth_token_ENV: GIGACHAT_GPT_TOKEN
+    api_type: gigachat
+    parallel: 1
+
+meta-llama-3-70b-instruct-gptq:
+    model_name: MaziyarPanahi/Meta-Llama-3-70B-Instruct-GPTQ
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+snorkel-mistral-pairrm-dpo:
+    model_name: snorkelai/Snorkel-Mistral-PairRM-DPO
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+sfr-iterative-dpo-llama-3-8b-r:
+    model_name: Salesforce/SFR-Iterative-DPO-LLaMA-3-8B-R
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+openchat-3.5-0106:
+    model_name: openchat/openchat-3.5-0106
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+mixtral-8x7b-instruct-v0.1:
+    model_name: LoneStriker/Mixtral-8x7B-Instruct-v0.1-HF
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 4
+
+neural-chat-7b-v3-3:
+    model_name: Intel/neural-chat-7b-v3-3
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+meta-llama-3-8b-instruct:
+    model_name: meta-llama/Meta-Llama-3-8B-Instruct
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+saiga_llama3_8b:
+    model_name: IlyaGusev/saiga_llama3_8b
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+hermes-2-pro-llama-3-8b:
+    model_name: NousResearch/Hermes-2-Pro-Llama-3-8B
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+dpopenhermes-7b:
+    model_name: openaccess-ai-collective/DPOpenHermes-7B
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+llama3-chatqa-1.5-8b:
+    model_name: nvidia/Llama3-ChatQA-1.5-8B
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+hermes-2-pro-mistral-7b:
+    model_name: NousResearch/Hermes-2-Pro-Mistral-7B
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+suzume-llama-3-8b-multilingual:
+    model_name: lightblue/suzume-llama-3-8B-multilingual
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+vikhr-7b-instruct_0.4:
+    model_name: Vikhrmodels/Vikhr-7B-instruct_0.4
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+vikhr-it-5.2-fp16-cp:
+    model_name: Vikhrmodels/it-5.2-fp16-cp
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    system_prompt: Ты — Вихрь, русскоязычный ассистент.
+    parallel: 6
+
+starling-lm-7b-beta:
+    model_name: Nexusflow/Starling-LM-7B-beta
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+c4ai-command-r-v01:
+    model_name: CohereForAI/c4ai-command-r-v01
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 6
+
+starcoder2-15b-instruct-v0.1:
+    model_name: bigcode/starcoder2-15b-instruct-v0.1
+    endpoints:
+        - api_base: http://localhost:8000/v1
+          api_key: token-abc123
+    api_type: openai
+    parallel: 3

src/gen/config/judge_config-ru.yaml

@@ -0,0 +1,35 @@
+name: judgment config file for Arena Hard
+
+bench_name: arena-hard-v0.1
+
+# Arena Hard default
+judge_model: gpt-4-1106-preview
+reference: False # Optional
+ref_model: null  
+
+baseline: True
+baseline_model: gpt-3.5-turbo-0125
+
+pairwise: True
+temperature: 0
+max_tokens: 4096
+
+regex_pattern: \[\[([AB<>=]+)\]\]
+
+system_prompt: "Пожалуйста, веди себя как беспристрастный судья и оцени качество ответов, предоставленных двумя AI ассистентами на пользовательский запрос, представленный ниже. Тебе будут даны ответы ассистента А и ассистента В. Твоя задача — оценить, чей ответ лучше.\n\nНачни свою оценку, сгенерировав собственный ответ на запрос. Ты должен предоставить свои ответы, прежде чем судить об ответах других AI.\n\nПри оценке ответов ассистентов сравни ответы обоих ассистентов со своим ответом. Ты должен идентифицировать и исправить любые ошибки или неточности.\n\nЗатем рассмотри, являются ли ответы ассистентов грамотными, полезными, релевантными и краткими. Грамотность означает, что ответ использует преимущественно русский язык и в нем отсутствуют языковые ошибки. Полезность означает, что ответ правильно реагирует на запрос или следует инструкциям. Обрати внимание, когда в запросе пользователя есть какая-либо неоднозначность или более одной интерпретации, полезнее и уместнее запрашивать уточнения или дополнительную информацию у пользователя, чем предоставлять ответ на основе предположений. Релевантность означает, что все части ответа тесно связаны или соотвествуют тому, что спрашивается. Краткость означает, что ответ ясен и не многословен или избыточен.\n\nЗатем рассмотри креативность и новизну ответов ассистентов, когда это необходимо. Наконец, определи любую отсутствующую важную информацию в ответах ассистентов, которую было бы полезно включить при ответе на пользовательский запрос.\n\nПосле предоставления твоего объяснения, ты должен выдать только один из следующих вариантов как твое окончательное решение с меткой:\n\n1. Ассистент A значительно лучше: [[A>>B]]\n2. Ассистент A немного лучше: [[A>B]]\n3. Ничья, примерно одинаково: [[A=B]]\n4. Ассистент B немного лучше: [[B>A]]\n5. Ассистент B значительно лучше: [[B>>A]]\n\nПример вывода: \"Мой окончательный вердикт — ничья: [[A=B]]\"."
+
+prompt_template: ["<|Запрос пользователя|>\n{question_1}\n\n<|Начало ответа ассистента A|>\n{answer_1}\n<|Конец ответа ассистента A|>\n\n<|Начало ответа ассистента B|>\n{answer_2}\n<|Конец ответа ассистента B|>"]
+
+# Add your model below for evaluation
+model_list:
+  - meta-llama-3-8b-instruct
+  - meta-llama-3-8b-instruct-ru-guided-2
+  - saiga_llama3_8b
+  - suzume-llama-3-8B-multilingual
+  - c4ai-command-r-v01
+  - starling-lm-7b-beta
+  - openchat-3.5-0106
+  - hermes-2-pro-llama-3-8b
+  - hermes-2-pro-mistral-7b
+  - starcoder2-15b-instruct-v0.1
+  - gpt-4-1106-preview

src/gen/config/judge_config.yaml

@@ -0,0 +1,40 @@
+name: judgment config file for Arena Hard
+
+bench_name: arena-hard-v0.1
+
+# Arena Hard default
+judge_model: gpt-4-1106-preview
+reference: False # Optional
+ref_model: null
+
+baseline: True
+baseline_model: gpt-3.5-turbo-0125
+
+pairwise: True
+temperature: 0
+max_tokens: 4096
+
+regex_pattern: \[\[([AB<>=]+)\]\]
+
+system_prompt: "Please act as an impartial judge and evaluate the quality of the responses provided by two AI assistants to the user prompt displayed below. You will be given assistant A's answer and assistant B's answer. Your job is to evaluate which assistant's answer is better.\n\nBegin your evaluation by describing the details that need to be taken into account when responding to this prompt. You must provide your ideas before judging any answers.\n\nWhen evaluating the assistants' answers, compare both assistants' answers with your ideas. You must identify and correct any mistakes or inaccurate information.\n\nThen consider if the assistant's answers are helpful, relevant, concise and linguistically acceptable. Helpful means the answer correctly responds to the prompt or follows the instructions. Note when user prompt has any ambiguity or more than one interpretation, it is more helpful and appropriate to ask for clarifications or more information from the user than providing an answer based on assumptions. Relevant means all parts of the response closely connect or are appropriate to what is being asked. Concise means the response is clear and not verbose or excessive. Linguistically acceptable means that the response is given mainly in Russian language and there are no grammatical errors in it.\n\nThen consider the creativity and novelty of the assistant's answers when needed. Finally, identify any missing important information in the assistants' answers that would be beneficial to include when responding to the user prompt.\n\nAfter providing your explanation, you must output only one of the following choices as your final verdict with a label:\n\n1. Assistant A is significantly better: [[A>>B]]\n2. Assistant A is slightly better: [[A>B]]\n3. Tie, relatively the same: [[A=B]]\n4. Assistant B is slightly better: [[B>A]]\n5. Assistant B is significantly better: [[B>>A]]\n\nExample output: \"My final verdict is tie: [[A=B]]\"."
+
+prompt_template: ["<|User Prompt|>\n{question_1}\n\n<|The Start of Assistant A's Answer|>\n{answer_1}\n<|The End of Assistant A's Answer|>\n\n<|The Start of Assistant B's Answer|>\n{answer_2}\n<|The End of Assistant B's Answer|>"]
+
+# Add your model below for evaluation
+model_list:
+  - meta-llama-3-8b-instruct
+  - saiga_llama3_8b
+  - suzume-llama-3-8b-multilingual
+  - yandex_gpt_pro
+  - c4ai-command-r-v01
+  - starling-lm-7b-beta
+  - openchat-3.5-0106
+  - snorkel-mistral-pairrm-dpo
+  - neural-chat-7b-v3-3
+  - gigachat_lite
+  - gigachat_pro
+  - vikhr-7b-instruct_0.4
+  - hermes-2-pro-llama-3-8b
+  - gpt-4-1106-preview
+  - llama3-chatqa-1.5-8b
+  - vikhr-it-5.1

src/gen/data/arena-hard-v0.1/model_answer/external/private/var/folders/ws/s9058_gn5cs181gs2_54lcvc0000gn/T/gradio/4a99fae57971a5f7e281df57ab8739fd979a9345/16.o1.csv

@@ -0,0 +1,11 @@
+Col1.Col2.Col3.Col4.Col5.Col6.Col7.Col8.Col9.Col10
+1.2.5.6.2.6.3.7.8.8
+10.10.10.7.8.3.8.9.4.8
+5.9.2.10.7.7.4.9.2.3
+4.8.2.9.8.7.6.6.9.4
+1.8.7.3.1.6.7.7.6.1
+9.9.6.2.1.5.5.2.5.5
+8.2.10.5.10.10.7.6.3.6
+6.1.8.3.3.4.7.7.8.5
+7.1.3.3.2.4.5.9.5.6
+4.1.4.4.6.1.2.6.9.2

src/gen/data/bootstrapping_results.jsonl

@@ -0,0 +1,100 @@
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.5644665503,"gigachat_lite":726.6208252619}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.0709454157,"gigachat_lite":738.5741612323}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.0434024226,"gigachat_lite":734.1011761886}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":860.399655762,"gigachat_lite":729.5571514643}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.1731508697,"gigachat_lite":728.758372467}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.5326400531,"gigachat_lite":733.7900136425}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.7819454641,"gigachat_lite":719.043685497}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":858.5219875589,"gigachat_lite":714.8370789545}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.4603125434,"gigachat_lite":725.8752720444}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.8350548067,"gigachat_lite":715.266084892}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":862.7609222876,"gigachat_lite":727.2017077065}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":854.2414273092,"gigachat_lite":739.3798608124}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":862.374147169,"gigachat_lite":719.6304899658}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":863.1792770928,"gigachat_lite":734.0546251412}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.2996605704,"gigachat_lite":718.4924449088}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":864.8988771163,"gigachat_lite":721.0729415472}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":867.0356240274,"gigachat_lite":738.5699274129}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":871.6157440982,"gigachat_lite":723.7105361329}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.9225322393,"gigachat_lite":728.2971721354}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":864.7557130348,"gigachat_lite":737.8461934603}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":853.284444198,"gigachat_lite":748.9971545908}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":851.7087385877,"gigachat_lite":713.1462726999}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":871.482425846,"gigachat_lite":720.2960317186}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.6122634027,"gigachat_lite":727.2517234335}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":852.7157509126,"gigachat_lite":694.2654473149}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.7938560994,"gigachat_lite":735.6639839406}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":874.1682886992,"gigachat_lite":730.5016731736}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.4589887037,"gigachat_lite":734.4551919945}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":850.0205093168,"gigachat_lite":728.8931636911}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":875.7282859976,"gigachat_lite":717.6726330463}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.3647024942,"gigachat_lite":733.3721052861}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":856.1797064852,"gigachat_lite":725.7981758416}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":867.6238850835,"gigachat_lite":731.0409312559}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.7097671655,"gigachat_lite":715.3647090465}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":874.4978660071,"gigachat_lite":737.7875979517}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.5650653089,"gigachat_lite":729.3512200797}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":890.8852955482,"gigachat_lite":715.9010959711}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.6426165155,"gigachat_lite":722.2116159282}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.3456423505,"gigachat_lite":724.6752254921}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.4854945486,"gigachat_lite":718.5749125859}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":880.1901418236,"gigachat_lite":723.0132896162}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":849.6103242372,"gigachat_lite":732.3587564613}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":871.0458800663,"gigachat_lite":740.6268654101}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":877.4244267245,"gigachat_lite":724.6297632896}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":875.3479511716,"gigachat_lite":743.701641735}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.1269918194,"gigachat_lite":723.5736702859}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.8015195801,"gigachat_lite":731.9752231934}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":868.2750694028,"gigachat_lite":722.3929635211}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":868.0957706924,"gigachat_lite":721.9705147906}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":870.6012679715,"gigachat_lite":738.9123529498}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":862.269673472,"gigachat_lite":733.7609432817}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":864.2488571071,"gigachat_lite":724.1850017217}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":874.1624601722,"gigachat_lite":727.8550112565}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":863.1194231025,"gigachat_lite":731.3315308989}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.1192986285,"gigachat_lite":722.5721295254}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":862.0030926827,"gigachat_lite":729.8940208849}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.5474187298,"gigachat_lite":735.9873637973}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":880.5566205251,"gigachat_lite":730.6501947523}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.7223684538,"gigachat_lite":702.8268457509}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":874.9512628918,"gigachat_lite":732.6491227137}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":858.7260910186,"gigachat_lite":736.225411771}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":871.4133525673,"gigachat_lite":745.6156113918}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.2715335516,"gigachat_lite":721.0912474577}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.3256361213,"gigachat_lite":736.2254117629}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.9022358038,"gigachat_lite":732.9674153867}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":867.5601382523,"gigachat_lite":723.0966793643}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":864.5272121008,"gigachat_lite":718.0704518208}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.7782194777,"gigachat_lite":722.2852812675}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.4086246736,"gigachat_lite":745.1185090985}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":870.0314924292,"gigachat_lite":736.9690722951}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.3587976891,"gigachat_lite":742.6306627437}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":851.5511568095,"gigachat_lite":733.1555506911}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":863.2094645624,"gigachat_lite":721.7491525609}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.0624318318,"gigachat_lite":723.0795022704}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":848.5397354473,"gigachat_lite":717.9478748234}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.9432204946,"gigachat_lite":726.703609728}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":861.2370229881,"gigachat_lite":725.3073844986}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":878.2964116149,"gigachat_lite":722.2116156669}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":857.9909782749,"gigachat_lite":720.1865370325}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":871.9069179589,"gigachat_lite":731.5240457448}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":860.2445059252,"gigachat_lite":737.0781670626}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":850.4012745111,"gigachat_lite":708.356058121}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":866.7922558028,"gigachat_lite":730.3511179714}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":862.2175409513,"gigachat_lite":727.5035049316}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":856.8494155845,"gigachat_lite":706.4191731996}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":856.4641060792,"gigachat_lite":734.2333848904}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":878.905415424,"gigachat_lite":736.5196621633}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":851.8853822745,"gigachat_lite":724.9647865416}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.2360763272,"gigachat_lite":718.7060814362}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":869.1579952553,"gigachat_lite":722.5615781913}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.2369472583,"gigachat_lite":731.6666527735}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":859.2009612357,"gigachat_lite":722.1914533305}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":876.2027799847,"gigachat_lite":719.1795542579}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":849.6362696273,"gigachat_lite":730.3223324585}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.1318475963,"gigachat_lite":724.1322488355}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":855.8791178271,"gigachat_lite":734.6332090556}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":873.3916447336,"gigachat_lite":716.1292305518}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":867.1797828548,"gigachat_lite":726.7846008592}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":865.1613697328,"gigachat_lite":717.027778133}
+{"gpt-3.5-turbo-0125":1000.0,"gigachat_pro":875.1689869302,"gigachat_lite":728.6562483681}

src/gen/gen_answer.py

@@ -0,0 +1,195 @@
+"""Generate answers using api endpoints.
+
+Usage:
+python gen_api_answer --parallel 32
+"""
+import argparse
+import json
+import os
+import time
+import concurrent.futures
+
+import tiktoken
+import shortuuid
+import tqdm
+
+from utils import (
+    load_questions,
+    load_model_answers,
+    make_config,
+    get_endpoint,
+    chat_completion_openai,
+    chat_completion_yandex,
+    chat_completion_gigachat,
+    chat_completion_anthropic,
+    chat_completion_openai_azure,
+    chat_completion_mistral,
+    chat_completion_gemini,
+    chat_completion_cohere,
+    reorg_answer_file,
+    OPENAI_MODEL_LIST,
+    temperature_config,
+)
+
+
+def get_answer(
+    question: dict, model: str, endpoint_info: dict, num_choices: int, max_tokens: int, temperature: float, answer_file: str, api_dict: dict
+):
+    if question["category"] in temperature_config:
+        temperature = temperature_config[question["category"]]
+
+    api_type = endpoint_info["api_type"]
+
+    conv = []
+
+    if "system_prompt" in endpoint_info.keys():
+        conv.append({"role": "system", "content": endpoint_info["system_prompt"]})
+    elif model in OPENAI_MODEL_LIST:
+        conv.append({"role": "system", "content": "You are a helpful assistant."})
+
+    encoding = tiktoken.encoding_for_model("gpt-3.5-turbo")
+    choices = []
+    for i in range(num_choices):
+        turns = []
+        for j in range(len(question["turns"])):
+            conv.append({"role": "user", "content": question["turns"][j]["content"]})
+            if api_type == "anthropic":
+                output = chat_completion_anthropic(model=endpoint_info["model_name"],
+                                                   messages=conv,
+                                                   temperature=temperature,
+                                                   max_tokens=max_tokens)
+            elif api_type == "mistral":
+                output = chat_completion_mistral(model=endpoint_info["model_name"],
+                                                 messages=conv,
+                                                 temperature=temperature,
+                                                 max_tokens=max_tokens)
+            elif api_type == "yandex":
+                output = chat_completion_yandex(model=endpoint_info["model_name"],
+                                                 messages=conv,
+                                                 temperature=temperature,
+                                                 max_tokens=max_tokens,
+                                                 api_dict=api_dict)
+            elif api_type == "gigachat":
+                output = chat_completion_gigachat(model=endpoint_info["model_name"],
+                                                 messages=conv,
+                                                 temperature=temperature,
+                                                 max_tokens=max_tokens,
+                                                 api_dict=api_dict)
+            elif api_type == "gemini":
+                output = chat_completion_gemini(model=endpoint_info["model_name"],
+                                                messages=question["turns"][j]["content"],
+                                                temperature=temperature,
+                                                max_tokens=max_tokens)
+            elif api_type == "azure":
+                output = chat_completion_openai_azure(model=endpoint_info["model_name"],
+                                                      messages=conv,
+                                                      temperature=temperature,
+                                                      max_tokens=max_tokens,
+                                                      api_dict=api_dict)
+            elif api_type == "cohere":
+                output = chat_completion_cohere(model=endpoint_info["model_name"],
+                                                messages=conv,
+                                                temperature=temperature,
+                                                max_tokens=max_tokens)
+            else:
+                output = chat_completion_openai(model=endpoint_info["model_name"],
+                                                messages=conv,
+                                                temperature=temperature,
+                                                max_tokens=max_tokens,
+                                                api_dict=api_dict)
+            conv.append({"role": "assistant", "content": output})
+
+            turns.append({"content": output, "token_len": len(encoding.encode(output))})
+        choices.append({"index": i, "turns": turns})
+
+    # Dump answers
+    ans = {
+        "question_id": question["question_id"],
+        "answer_id": shortuuid.uuid(),
+        "model_id": model,
+        "choices": choices,
+        "tstamp": time.time(),
+    }
+
+    os.makedirs(os.path.dirname(answer_file), exist_ok=True)
+    with open(answer_file, "a") as fout:
+        fout.write(json.dumps(ans) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--setting-file", type=str, default="config/gen_answer_config.yaml"
+    )
+    parser.add_argument(
+        "--endpoint-file", type=str, default="config/api_config.yaml"
+    )
+    args = parser.parse_args()
+
+    settings = make_config(args.setting_file)
+    endpoint_list = make_config(args.endpoint_file)
+
+    existing_answer = load_model_answers(os.path.join("data", settings["bench_name"], "model_answer"))
+
+    print(settings)
+
+    for model in settings["model_list"]:
+        assert model in endpoint_list
+        endpoint_info = endpoint_list[model]
+
+        question_file = os.path.join("data", settings["bench_name"], "question.jsonl")
+        questions = load_questions(question_file)
+
+        answer_file = os.path.join("data", settings["bench_name"], "model_answer", f"{model}.jsonl")
+        print(f"Output to {answer_file}")
+
+        if "parallel" in endpoint_info:
+            parallel = endpoint_info["parallel"]
+        else:
+            parallel = 1
+
+        # We want to maximizes the number of tokens generate per answer: max_tokens = specified token # - input tokens #
+        if "tokenizer" in endpoint_info:
+            question_list = [question["turns"][0]["content"] for question in questions]
+            if model in OPENAI_MODEL_LIST:
+                tokenizer = tiktoken.encoding_for_model(endpoint_info["model_name"])
+                tokens = [tokenizer.encode(prompt) for prompt in question_list]
+                max_tokens = [(settings["max_tokens"] - len(token) - 100) for token in tokens]
+            else:
+                from transformers import AutoTokenizer
+
+                os.environ["TOKENIZERS_PARALLELISM"] = "false"
+                tokenizer = AutoTokenizer.from_pretrained(endpoint_info["tokenizer"])
+
+                tokens = tokenizer(question_list)
+                max_tokens = [(settings["max_tokens"] - len(prompt) - 300) for prompt in tokens["input_ids"]]
+        else:
+            max_tokens = [settings["max_tokens"]] * len(questions)
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=parallel) as executor:
+            futures = []
+            count = 0
+            for index, question in enumerate(questions):
+                if model in existing_answer and question["question_id"] in existing_answer[model]:
+                    count += 1
+                    continue
+                future = executor.submit(
+                    get_answer,
+                    question,
+                    model,
+                    endpoint_info,
+                    settings["num_choices"],
+                    max_tokens[index],
+                    settings["temperature"],
+                    answer_file,
+                    get_endpoint(endpoint_info["endpoints"]),
+                )
+                futures.append(future)
+            if count > 0:
+                print(f"{count} number of existing answers")
+            for future in tqdm.tqdm(
+                concurrent.futures.as_completed(futures), total=len(futures)
+            ):
+                future.result()
+
+        reorg_answer_file(answer_file)

src/gen/gen_judgment.py

@@ -0,0 +1,220 @@
+import argparse
+import concurrent.futures
+import glob
+import json
+import os
+import re
+
+import huggingface_hub
+from tqdm import tqdm
+from utils import (
+    chat_completion_anthropic,
+    chat_completion_openai,
+    chat_completion_openai_azure,
+    get_endpoint,
+    load_model_answers,
+    load_questions,
+    make_config,
+)
+
+
+def get_score(judgment, pattern, pairwise=True):
+    matches = pattern.findall(judgment)
+    matches = [m for m in matches if m != ""]
+    if len(set(matches)) == 0:
+        return None, True
+    elif len(set(matches)) == 1:
+        if pairwise:
+            return matches[0].strip("\n"), False
+        return int(matches[0])
+    else:
+        return None, False
+
+
+# get answer from model
+def get_answer(model, conv, temperature, max_tokens, endpoint_dict=None):
+    api_dict = get_endpoint(endpoint_dict["endpoints"])
+
+    if endpoint_dict["api_type"] == "anthropic":
+        output = chat_completion_anthropic(model, conv, temperature, max_tokens)
+    elif endpoint_dict["api_type"] == "azure":
+        output = chat_completion_openai_azure(model, conv, temperature, max_tokens, api_dict)
+    else:
+        output = chat_completion_openai(model, conv, temperature, max_tokens, api_dict)
+    return output
+
+
+def judgment(**args):
+    question = args["question"]
+    answer = args["answer"]
+    reference = args["reference"]
+    baseline = args["baseline_answer"]
+    configs = args["configs"]
+    output_file = args["output_file"]
+    model = configs["judge_model"]
+
+    num_games = 2 if configs["pairwise"] else 1
+
+    output = {
+        "question_id":question["question_id"],
+        "model":answer["model_id"],
+        "judge": model,
+        "games":[]
+        }
+
+    for game in range(num_games):
+        conv = [{"role": "system", "content": configs["system_prompt"]}]
+
+        for template in configs["prompt_template"]:
+            prompt_args = {}
+
+            for i, turn in enumerate(question["turns"]):
+                prompt_args[f"question_{i+1}"] = turn["content"]
+            base = 1
+
+            if baseline:
+                if game % 2 == 1: # swap position
+                    temp = baseline
+                    baseline = answer
+                    answer = temp
+
+                for i, turn in enumerate(baseline["choices"][0]["turns"]):
+                    prompt_args[f"answer_{i+1}"] = turn["content"]
+                    base += 1
+            if answer:
+                for i, turn in enumerate(answer["choices"][0]["turns"]):
+                    prompt_args[f"answer_{i+base}"] = turn["content"]
+
+            if reference:
+                for j, ref_answer in enumerate(reference):
+                    for i, turn in enumerate(ref_answer["choices"][0]["turns"]):
+                        prompt_args[f"ref_answer_{i+j+1}"] = turn["content"]
+
+            user_prompt = template.format(**prompt_args)
+            conv.append({"role": "user", "content": user_prompt})
+
+        judgment = ""
+        for _ in range(2):
+            new_judgment = get_answer(
+                model,
+                conv,
+                configs["temperature"],
+                configs["max_tokens"],
+                args["endpoint_dict"],
+            )
+
+            judgment += ("\n" + new_judgment)
+
+            score, try_again = get_score(judgment, args["regex_pattern"])
+
+            conv.append({"role": "assistant", "content": new_judgment})
+
+            if not try_again:
+                break
+
+            conv.append({"role": "user", "content": "continue your judgment and finish by outputting a final verdict label"})
+
+        result = {
+            "user_prompt": conv[1]["content"],
+            "judgment": judgment,
+            "score":score
+        }
+        output["games"].append(result)
+
+    with open(output_file, "a") as f:
+        f.write(json.dumps(output, ensure_ascii=False) + "\n")
+    huggingface_hub.HfApi().upload_file(output_file, path_in_repo=f'model_judgment/{configs['judge_model']}/{output_file.split('/')[-1]}', repo_id='Vikhrmodels/openbench-eval', repo_type='dataset')
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--setting-file", type=str, default="./config/judge_config.yaml")
+    parser.add_argument("--endpoint-file", type=str, default="./config/api_config.yaml")
+    args = parser.parse_args()
+    print(args)
+
+    configs = make_config(args.setting_file)
+    endpoint_list = make_config(args.endpoint_file)
+
+    print(f'judge model: {configs["judge_model"]}, baseline: {configs["baseline"]}, baseline model: {configs["baseline_model"]}, reference: {configs["reference"]}, '
+          + f'reference models: {configs["ref_model"]}, temperature: {configs["temperature"]}, max tokens: {configs["max_tokens"]}, pairwise: {configs["pairwise"]}')
+
+    if configs["regex_pattern"]:
+        pattern = re.compile(configs["regex_pattern"])
+
+    question_file = os.path.join("./data", configs["bench_name"], "question.jsonl")
+    external_dir = os.path.join("./data", configs["bench_name"], "model_answer/external")
+    internal_dir = os.path.join("./data", configs["bench_name"], "model_answer/internal")
+    ref_answer_dir = os.path.join("data", configs["bench_name"], "reference_answer")
+
+    questions = load_questions(question_file)
+    model_answers_external = load_model_answers(external_dir)
+    model_answers_internal = load_model_answers(internal_dir)
+    
+    # internal has priority
+    model_answers = {**model_answers_external, **model_answers_internal}
+
+    # if user choose a set of models, only judge those models
+    models = [model.split('/')[-1].split('.')[0] for model in glob.glob('./data/arena-hard-v0.1/model_answer/external/*.jsonl')]
+
+    ref_answers = None
+    if configs["reference"]:
+        ref_answers = load_model_answers(ref_answer_dir)
+        ref_answers = [ref_answers[model] for model in configs["ref_model"]]
+
+    output_files = {}
+    output_dir = f"data/{configs['bench_name']}/model_judgment/{configs['judge_model']}"
+    for model in models:
+        output_files[model] = os.path.join(
+            output_dir,
+            f"{model}.jsonl",
+        )
+
+    for output_file in output_files.values():
+        os.makedirs(os.path.dirname(output_file), exist_ok=True)
+
+    existing_judgments = load_model_answers(output_dir)
+
+    endpoint_info = endpoint_list[configs["judge_model"]]
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=endpoint_info["parallel"]) as executor:
+        futures = []
+        for model in models:
+            count = 0
+            for question in questions[:2]:
+                question_id = question["question_id"]
+
+                kwargs = {}
+                kwargs["question"] = question
+                if model in model_answers and question_id not in model_answers[model]:
+                    print(f"Warning: {model} answer to {question['question_id']} cannot be found.")
+                    continue
+
+                if model in existing_judgments and question_id in existing_judgments[model]:
+                    count += 1
+                    continue
+
+                kwargs["answer"] = model_answers[model][question_id]
+                if ref_answers:
+                    kwargs["reference"] = [ref_answer[question_id] for ref_answer in ref_answers]
+                    assert len(kwargs["reference"]) == len(configs["ref_model"])
+                else:
+                    kwargs["reference"] = None
+                if configs["baseline"]:
+                    kwargs["baseline_answer"] = model_answers[configs["baseline_model"]][question_id]
+                else:
+                    kwargs["baseline_answer"] = None
+                kwargs["configs"] = configs
+                kwargs["endpoint_dict"] = endpoint_info
+                kwargs["output_file"] = output_files[model]
+                kwargs["regex_pattern"] = pattern
+                future = executor.submit(judgment, **kwargs)
+                futures.append(future)
+
+            if count > 0:
+                print(f"{count} number of existing judgments")
+
+        for future in tqdm(
+            concurrent.futures.as_completed(futures), total=len(futures)
+        ):
+            future.result()

src/gen/show_result.py

@@ -0,0 +1,258 @@
+import pandas as pd
+import numpy as np
+import plotly.express as px
+
+import tiktoken
+import datetime
+import argparse
+import os
+import math
+
+from glob import glob
+from tqdm import tqdm
+
+from sklearn.linear_model import LogisticRegression
+from collections import defaultdict
+from utils import load_model_answers
+
+def compute_mle_elo(df, SCALE=400, BASE=10, INIT_RATING=1000):
+    models = pd.concat([df["model_a"], df["model_b"]]).unique()
+    models = pd.Series(np.arange(len(models)), index=models)
+
+    # duplicate battles
+    df = pd.concat([df, df], ignore_index=True)
+    p = len(models.index)
+    n = df.shape[0]
+
+    X = np.zeros([n, p])
+    X[np.arange(n), models[df["model_a"]]] = +math.log(BASE)
+    X[np.arange(n), models[df["model_b"]]] = -math.log(BASE)
+
+    # one A win => two A win
+    Y = np.zeros(n)
+    Y[df["winner"] == "model_a"] = 1.0
+
+    # one tie => one A win + one B win
+    # find tie + tie (both bad) index
+    tie_idx = (df["winner"] == "tie") | (df["winner"] == "tie (bothbad)")
+    tie_idx[len(tie_idx)//2:] = False
+    Y[tie_idx] = 1.0
+
+    lr = LogisticRegression(fit_intercept=False, penalty=None, tol=1e-8)
+    lr.fit(X,Y)
+
+    elo_scores = SCALE * lr.coef_[0] + INIT_RATING
+
+    # set anchor as gpt-3.5-turbo-0125 = 1000
+    if "gpt-3.5-turbo-0125" in models.index:
+        elo_scores += 1000 - elo_scores[models["gpt-3.5-turbo-0125"]]
+    return pd.Series(elo_scores, index = models.index).sort_values(ascending=False)
+
+
+def get_bootstrap_result(battles, func_compute_elo, num_round):
+    rows = []
+    for i in tqdm(range(num_round), desc="bootstrap"):
+        rows.append(func_compute_elo(battles.sample(frac=1.0, replace=True)))
+    df = pd.DataFrame(rows)
+    return df[df.median().sort_values(ascending=False).index]
+
+
+def preety_print_two_ratings(ratings_1, ratings_2, column_names):
+    df = pd.DataFrame([
+        [n, ratings_1[n], ratings_2[n]] for n in ratings_1.keys()
+    ], columns=["Model", column_names[0], column_names[1]]).sort_values(column_names[0], ascending=False).reset_index(drop=True)
+    df[column_names[0]] = (df[column_names[0]] + 0.5).astype(int)
+    df[column_names[1]] = (df[column_names[1]] + 0.5).astype(int)
+    df.index = df.index + 1
+    return df
+
+
+def visualize_bootstrap_scores(df, title):
+    bars = pd.DataFrame(dict(
+        lower = df.quantile(.025),
+        rating = df.quantile(.5),
+        upper = df.quantile(.975))).reset_index(names="model").sort_values("rating", ascending=False)
+    bars['error_y'] = bars['upper'] - bars["rating"]
+    bars['error_y_minus'] = bars['rating'] - bars["lower"]
+    bars['rating_rounded'] = np.round(bars['rating'], 2)
+    fig = px.scatter(bars, x="model", y="rating", error_y="error_y",
+                     error_y_minus="error_y_minus", text="rating_rounded",
+                     title=title)
+    fig.update_layout(xaxis_title="Model", yaxis_title="Rating",
+                      height=600)
+    return fig
+
+
+def predict_win_rate(elo_ratings, SCALE=400, BASE=10, INIT_RATING=1000):
+    names = sorted(list(elo_ratings.keys()))
+    wins = defaultdict(lambda: defaultdict(lambda: 0))
+    for a in names:
+        for b in names:
+            ea = 1 / (1 + BASE ** ((elo_ratings[b] - elo_ratings[a]) / SCALE))
+            wins[a][b] = ea
+            wins[b][a] = 1 - ea
+
+    data = {
+        a: [wins[a][b] if a != b else np.NAN for b in names]
+        for a in names
+    }
+
+    df = pd.DataFrame(data, index=names)
+    df.index.name = "model_a"
+    df.columns.name = "model_b"
+    return df.T
+
+
+def get_win_rate_column(df, column, baseline="gpt-3.5-turbo-0125"):
+    to_dict = df[["model", column]].set_index("model").to_dict()[column]
+    win_rate_table = predict_win_rate(to_dict)
+    return win_rate_table[baseline].fillna(0.5).apply(lambda x: round(x * 100, 2))
+
+
+def get_battles_from_judgment(judge_name, first_game_only=False, WEIGHT=3):
+    arena_hard_battles = pd.DataFrame()
+
+    print("Turning judgment results into battles...")
+
+    directory = f"data/arena-hard-v0.1/model_judgement/{judge_name}"
+    assert os.path.exists(directory)
+    for file in tqdm(glob(f"{directory}/*jsonl")):
+        df = pd.read_json(file, lines=True)
+
+        for _, row in df.iterrows():
+            # game 1
+            output = {"question_id": row["question_id"],
+                    "model_a": "gpt-3.5-turbo-0125",
+                    "model_b": row["model"]}
+
+            game = row["games"][0]
+
+            weight = 1
+            if game["score"] == "A=B":
+                output["winner"] = "tie"
+            elif game["score"] == "A>B":
+                output["winner"] = "model_a"
+            elif game["score"] == "A>>B":
+                output["winner"] = "model_a"
+                weight = WEIGHT
+            elif game["score"] == "B>A":
+                output["winner"] = "model_b"
+            elif game["score"] == "B>>A":
+                output["winner"] = "model_b"
+                weight = WEIGHT
+            else:
+                weight = 0
+
+            if weight:
+                arena_hard_battles = pd.concat([arena_hard_battles, pd.DataFrame([output] * weight)])
+
+            if not first_game_only:
+                # game 2
+                output = {"question_id": row["question_id"],
+                        "model_a": "gpt-3.5-turbo-0125",
+                        "model_b": row["model"]}
+
+                game = row["games"][1]
+
+                weight = 1
+                if game["score"] == "A=B":
+                    output["winner"] = "tie"
+                elif game["score"] == "A>B":
+                    output["winner"] = "model_b"
+                elif game["score"] == "A>>B":
+                    output["winner"] = "model_b"
+                    weight = WEIGHT
+                elif game["score"] == "B>A":
+                    output["winner"] = "model_a"
+                elif game["score"] == "B>>A":
+                    output["winner"] = "model_a"
+                    weight = WEIGHT
+                else:
+                    weight = 0
+
+                if weight:
+                    arena_hard_battles = pd.concat([arena_hard_battles, pd.DataFrame([output] * weight)])
+    arena_hard_battles.to_json("data/arena_hard_battles.jsonl", lines=True, orient="records")
+    return arena_hard_battles
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--bench-name", type=str, default="arena-hard-v0.1")
+    parser.add_argument("--judge-name", type=str, default="gpt-4-1106-preview")
+    parser.add_argument("--baseline", type=str, default="gpt-3.5-turbo-0125")
+    parser.add_argument("--load-battles", action="store_true")
+    parser.add_argument("--load-bootstrap", action="store_true")
+    parser.add_argument("--show-elo", action="store_true")
+    parser.add_argument("--weight", type=int, default=3)
+    parser.add_argument("--num-rounds", type=int, default=100)
+    parser.add_argument("--output", action="store_true")
+    parser.add_argument("--first-game-only", action="store_true")
+    args = parser.parse_args()
+    print(args)
+    assert not args.load_bootstrap or (args.load_battles and args.load_bootstrap), "If loading prexisting bootstrapping data, you must also load preexisting battles."
+
+    answer_dir = os.path.join("data", args.bench_name, "model_answer/external")
+    model_answers = load_model_answers(answer_dir)
+
+    if args.load_battles:
+        assert os.path.exists("data/arena_hard_battles.jsonl")
+        battles = pd.read_json("data/arena_hard_battles.jsonl", lines=True)
+    else:
+        battles = get_battles_from_judgment(args.judge_name, args.first_game_only, args.weight)
+
+    bootstrap_online_elo = compute_mle_elo(battles)
+
+
+    if args.load_bootstrap:
+        bootstrap_elo_lu = pd.read_json("data/bootstrapping_results.jsonl", lines=True)
+    else:
+        np.random.seed(42)
+        bootstrap_elo_lu = get_bootstrap_result(battles, compute_mle_elo, args.num_rounds)
+        bootstrap_elo_lu.to_json("data/bootstrapping_results.jsonl", lines=True, orient="records")
+
+    stats = pd.DataFrame()
+    stats["results"] = None
+    stats["results"] = stats['results'].astype('object')
+
+    for i, model in enumerate(bootstrap_online_elo.index):
+        assert model in bootstrap_elo_lu.columns
+
+        stats.at[i, "model"] = model
+        stats.at[i, "score"] = bootstrap_online_elo[model]
+        stats.at[i, "lower"] = np.percentile(bootstrap_elo_lu[model], 2.5)
+        stats.at[i, "upper"] = np.percentile(bootstrap_elo_lu[model], 97.5)
+
+        length = 0
+        if model in model_answers:
+            for _, row in model_answers[model].items():
+                turn = row["choices"][0]["turns"][0]
+                length += turn["token_len"]
+            length /= len(model_answers[model])
+
+        stats.at[i, "avg_tokens"] = int(length)
+        stats.at[i, "results"] = bootstrap_elo_lu[model].tolist()
+
+    if not args.show_elo:
+        stats.sort_values(by="model", inplace=True)
+        stats["score"] = get_win_rate_column(stats, "score", args.baseline).tolist()
+        stats["lower"] = get_win_rate_column(stats, "lower", args.baseline).tolist()
+        stats["upper"] = get_win_rate_column(stats, "upper", args.baseline).tolist()
+        decimal = 1
+    else:
+        decimal = 0
+        stats = stats.astype({"score" : int, "lower" : int, "upper" : int})
+
+    stats.sort_values(by="score", ascending=False, inplace=True)
+    for _, row in stats.iterrows():
+        interval = str((round(row['lower'] - row['score'], decimal), round(row['upper'] - row['score'], decimal)))
+        print(f"{row['model'] : <30} | score: {round(row['score'], decimal) : ^5} | 95% CI: {interval : ^12} | average #tokens: {int(row['avg_tokens'])}")
+
+    if args.output:
+        cur_date = datetime.datetime.now()
+        date_str = cur_date.strftime("%Y%m%d")
+        stats.to_json(f"arena_hard_leaderboard_{date_str}.json", orient="records", indent=4)
+        import huggingface_hub
+        huggingface_hub.HfApi().upload_file(path_or_fileobj=f"arena_hard_leaderboard_{date_str}.json",path_in_repo='evals/upd.json',
+                                          repo_id='Vikhrmodels/openbench-eval',
+                                          repo_type='dataset')

src/gen/utils.py

@@ -0,0 +1,394 @@
+import json
+import os
+import random
+import time
+from glob import glob
+
+import yaml
+
+# API setting constants
+API_MAX_RETRY = 16
+API_RETRY_SLEEP = 10
+API_ERROR_OUTPUT = "$ERROR$"
+
+
+OPENAI_MODEL_LIST = (
+    "gpt-3.5-turbo",
+    "gpt-3.5-turbo-0301",
+    "gpt-3.5-turbo-0613",
+    "gpt-3.5-turbo-0613-verbose",
+    "gpt-3.5-turbo-1106",
+    "gpt-3.5-turbo-0125",
+    "gpt-4",
+    "gpt-4-0314",
+    "gpt-4-0613",
+    "gpt-4-turbo",
+    "gpt-4-1106-preview",
+    "gpt-4-0125-preview",
+)
+
+
+temperature_config = {
+    "writing": 0.7,
+    "roleplay": 0.7,
+    "extraction": 0.0,
+    "math": 0.0,
+    "coding": 0.0,
+    "reasoning": 0.0,
+    "stem": 0.1,
+    "humanities": 0.1,
+}
+
+
+def load_questions(question_file: str):
+    """Load questions from a file."""
+    questions = []
+    with open(question_file, "r") as ques_file:
+        for line in ques_file:
+            if line:
+                questions.append(json.loads(line))
+    return questions
+
+
+def load_model_answers(answer_dir: str):
+    """Load model answers.
+
+    The return value is a python dict of type:
+    Dict[model_name: str -> Dict[question_id: int -> answer: dict]]
+    """
+    filenames = glob(os.path.join(answer_dir, "*.jsonl"))
+    filenames.sort()
+    model_answers = {}
+
+    for filename in filenames:
+        model_name = os.path.basename(filename)[:-6]
+        answer = {}
+        with open(filename) as fin:
+            for line in fin:
+                line = json.loads(line)
+                answer[line["question_id"]] = line
+        model_answers[model_name] = answer
+
+    return model_answers
+
+
+def get_endpoint(endpoint_list):
+    if endpoint_list is None:
+        return None
+    assert endpoint_list is not None
+    # randomly pick one
+    api_dict = random.choices(
+        endpoint_list
+    )[0]
+    return api_dict
+
+
+# load config args from config yaml files
+def make_config(config_file: str) -> dict:
+    config_kwargs = {}
+    with open(config_file, "r") as f:
+        config_kwargs = yaml.load(f, Loader=yaml.SafeLoader)
+
+    return config_kwargs
+
+def chat_completion_gigachat(model, messages, temperature, max_tokens, api_dict=None):
+    from gigachat import GigaChat
+    from gigachat.models import Chat, Messages
+    assert api_dict is not None, "no api settings provided!"
+    auth_token = api_dict.get("auth_token", os.environ.get(api_dict["auth_token"], ""))
+    client = GigaChat(credentials=auth_token, model=model, verify_ssl_certs=False)
+    temperature = max(temperature, 0.001)
+
+    messages = [Messages.parse_obj(m) for m in messages]
+    chat = Chat(messages=messages, max_tokens=max_tokens, temperature=temperature)
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            output = client.chat(chat)
+            output = output.choices[0].message.content
+            break
+        # Don't know other errors
+        except Exception as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+
+    return output
+
+def chat_completion_yandex(model, messages, temperature, max_tokens, api_dict=None):
+    from yandex_gpt import YandexGPT, YandexGPTConfigManagerForIAMToken
+    assert api_dict is not None, "no api settings provided!"
+    iam_token = api_dict.get("iam_token", os.environ.get(api_dict["iam_token_ENV"], ""))
+    config = YandexGPTConfigManagerForIAMToken(
+        model_type=model,
+        catalog_id=api_dict["catalog_id"],
+        iam_token=iam_token
+    )
+    client = YandexGPT(config_manager=config)
+
+    messages = [{"role": m["role"], "text": m["content"]} for m in messages]
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            output = client.get_sync_completion(
+                messages=messages,
+                temperature=temperature,
+                max_tokens=max_tokens,
+            )
+            break
+        # Don't know other errors
+        except Exception as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+
+    return output
+
+
+def chat_completion_openai(model, messages, temperature, max_tokens, api_dict=None):
+    import openai
+    api_key = api_dict.get("api_key", os.environ.get(api_dict["api_key_ENV"], ""))
+    if api_dict:
+        client = openai.OpenAI(
+            base_url=api_dict["api_base"],
+            api_key=api_key,
+        )
+    else:
+        client = openai.OpenAI()
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            # print(messages)
+            completion = client.chat.completions.create(
+                model=model,
+                messages=messages,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                stop=["</s>", "<eos>", "<|eot_id|>"]
+                )
+            output = completion.choices[0].message.content
+            break
+        except openai.RateLimitError as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+        except openai.BadRequestError as e:
+            print(messages)
+            print(type(e), e)
+        except KeyError:
+            print(type(e), e)
+            break
+
+    return output
+
+
+def chat_completion_openai_azure(model, messages, temperature, max_tokens, api_dict=None):
+    import openai
+    from openai import AzureOpenAI
+
+    api_base = api_dict["api_base"]
+    api_key = api_dict.get("api_key", os.environ.get(api_dict["api_key_ENV"], ""))
+    client = AzureOpenAI(
+        azure_endpoint = api_base,
+        api_key= api_key,
+        api_version=api_dict["api_version"],
+        timeout=240,
+        max_retries=2
+    )
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            response = client.chat.completions.create(
+                model=model,
+                messages=messages,
+                n=1,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                seed=42,
+            )
+            output = response.choices[0].message.content
+            break
+        except openai.RateLimitError as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+        except openai.BadRequestError as e:
+            print(type(e), e)
+            break
+        except KeyError:
+            print(type(e), e)
+            break
+
+    return output
+
+
+def chat_completion_anthropic(model, messages, temperature, max_tokens, api_dict=None):
+    import anthropic
+
+    if api_dict:
+        api_key = api_dict.get("api_key", os.environ.get(api_dict["api_key_ENV"], ""))
+    else:
+        api_key = os.environ["ANTHROPIC_API_KEY"]
+
+    sys_msg = ""
+    if messages[0]["role"] == "system":
+        sys_msg = messages[0]["content"]
+        messages = messages[1:]
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            # print(sys_msg)
+            c = anthropic.Anthropic(api_key=api_key)
+            response = c.messages.create(
+                model=model,
+                messages=messages,
+                stop_sequences=[anthropic.HUMAN_PROMPT],
+                max_tokens=max_tokens,
+                temperature=temperature,
+                system=sys_msg
+            )
+            output = response.content[0].text
+            break
+        except anthropic.APIError as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+    return output
+
+
+def chat_completion_mistral(model, messages, temperature, max_tokens):
+    from mistralai.client import MistralClient
+    from mistralai.exceptions import MistralException
+    from mistralai.models.chat_completion import ChatMessage
+
+    api_key = os.environ["MISTRAL_API_KEY"]
+    client = MistralClient(api_key=api_key)
+
+    prompts = [ChatMessage(role=message["role"], content=message["content"]) for message in messages]
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            chat_response = client.chat(
+                model=model,
+                messages=prompts,
+                temperature=temperature,
+                max_tokens=max_tokens,
+            )
+            output = chat_response.choices[0].message.content
+            break
+        except MistralException as e:
+            print(type(e), e)
+            break
+
+    return output
+
+
+def chat_completion_gemini(model, messages, temperature, max_tokens):
+    import google.generativeai as genai
+    genai.configure(api_key=os.environ["GEMINI_API_KEY"])
+
+    safety_settings = [
+        {
+            "category": "HARM_CATEGORY_HARASSMENT",
+            "threshold": "BLOCK_NONE"
+        },
+        {
+            "category": "HARM_CATEGORY_HATE_SPEECH",
+            "threshold": "BLOCK_NONE"
+        },
+        {
+            "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
+            "threshold": "BLOCK_NONE"
+        },
+        {
+            "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
+            "threshold": "BLOCK_NONE"
+        },
+    ]
+
+    # Set up the model
+    generation_config = {
+        "temperature": temperature,
+        "top_p": 1,
+        "top_k": 1,
+        "max_output_tokens": max_tokens,
+    }
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            gemini = genai.GenerativeModel(
+                model_name=model,
+                generation_config=generation_config,
+                safety_settings=safety_settings)
+
+            convo = gemini.start_chat(history=[])
+
+            convo.send_message(messages)
+            output = convo.last.text
+            break
+        except genai.types.generation_types.StopCandidateException as e:
+            print(type(e), e)
+            break
+        except Exception as e:
+            print(type(e), e)
+            time.sleep(API_RETRY_SLEEP)
+
+    return output
+
+
+def chat_completion_cohere(model, messages, temperature, max_tokens):
+    import cohere
+
+    co = cohere.Client(os.environ["COHERE_API_KEY"])
+    assert len(messages) > 0
+
+    template_map = {"system":"SYSTEM",
+                    "assistant":"CHATBOT",
+                    "user":"USER"}
+
+    assert messages[-1]["role"] == "user"
+    prompt = messages[-1]["content"]
+
+    if len(messages) > 1:
+        history = []
+        for message in messages[:-1]:
+            history.append({"role":template_map[message["role"]], "message":message["content"]})
+    else:
+        history = None
+
+    output = API_ERROR_OUTPUT
+    for _ in range(API_MAX_RETRY):
+        try:
+            response = co.chat(
+                message=prompt,
+                model=model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                chat_history=history,
+            )
+            output = response.text
+            break
+        except cohere.core.api_error.ApiError as e:
+            print(type(e), e)
+            raise
+        except Exception as e:
+            print(type(e), e)
+            break
+
+    return output
+
+
+def reorg_answer_file(answer_file):
+    """Sort by question id and de-duplication"""
+    answers = {}
+    with open(answer_file, "r") as fin:
+        for l in fin:
+            qid = json.loads(l)["question_id"]
+            answers[qid] = l
+
+    qids = sorted(list(answers.keys()))
+    with open(answer_file, "w") as fout:
+        for qid in qids:
+            fout.write(answers[qid])

src/leaderboard/filter_models.py

@@ -0,0 +1,174 @@
+from src.display.formatting import model_hyperlink
+from src.display.utils import AutoEvalColumn
+
+
+# Models which have been flagged by users as being problematic for a reason or another
+# (Model name to forum discussion link)
+FLAGGED_MODELS = {
+    "merged": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Voicelab/trurl-2-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/202",
+    "deepnight-research/llama-2-70B-inst": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/207",
+    "Aspik101/trurl-2-13b-pl-instruct_unload": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/213",
+    "Fredithefish/ReasonixPajama-3B-HF": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/236",
+    "TigerResearch/tigerbot-7b-sft-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/237",
+    "gaodrew/gaodrew-gorgonzola-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/215",
+    "AIDC-ai-business/Marcoroni-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+    "AIDC-ai-business/Marcoroni-13B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+    "AIDC-ai-business/Marcoroni-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+    "fblgit/una-xaberius-34b-v1beta": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/444",
+    "jan-hq/trinity-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "rwitz2/go-bruins-v2.1.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "rwitz2/go-bruins-v2.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "GreenNode/GreenNodeLM-v3olet-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "GreenNode/GreenNodeLM-7B-v4leo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "GreenNode/LeoScorpius-GreenNode-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "viethq188/LeoScorpius-7B-Chat-DPO": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "GreenNode/GreenNodeLM-7B-v2leo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "janai-hq/trinity-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "ignos/LeoScorpius-GreenNode-Alpaca-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "fblgit/una-cybertron-7b-v3-OMA": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "mncai/mistral-7b-dpo-merge-v1.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "mncai/mistral-7b-dpo-v6": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "Toten5/LeoScorpius-GreenNode-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "GreenNode/GreenNodeLM-7B-v1olet": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "quantumaikr/quantum-dpo-v0.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "quantumaikr/quantum-v0.01": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "quantumaikr/quantum-trinity-v0.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "mncai/mistral-7b-dpo-v5": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "cookinai/BruinHermes": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "jan-ai/Pandora-10.7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "v1olet/v1olet_marcoroni-go-bruins-merge-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "v1olet/v1olet_merged_dpo_7B_v3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "rwitz2/pee": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "zyh3826 / GML-Mistral-merged-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/503",
+    "dillfrescott/trinity-medium": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
+    "udkai/Garrulus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/526",
+    "dfurman/GarrulusMarcoro-7B-v0.1": "https://huggingface.co/dfurman/GarrulusMarcoro-7B-v0.1/discussions/1",
+    "eren23/slerp-test-turdus-beagle": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "abideen/NexoNimbus-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "alnrg2arg/test2_3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "nfaheem/Marcoroni-7b-DPO-Merge": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "CultriX/MergeTrix-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    "liminerity/Blur-7b-v1.21": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
+    # Merges not indicated
+    "gagan3012/MetaModelv2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "gagan3012/MetaModelv3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "kyujinpy/Sakura-SOLRCA-Math-Instruct-DPO-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "kyujinpy/Sakura-SOLAR-Instruct-DPO-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "kyujinpy/Sakura-SOLRCA-Math-Instruct-DPO-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "kyujinpy/Sakura-SOLRCA-Instruct-DPO": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "fblgit/LUNA-SOLARkrautLM-Instruct": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "perlthoughts/Marcoroni-8x7B-v3-MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "rwitz/go-bruins-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "rwitz/go-bruins": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Walmart-the-bag/Solar-10.7B-Cato": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "aqweteddy/mistral_tv-neural-marconroni": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "NExtNewChattingAI/shark_tank_ai_7_b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Q-bert/MetaMath-Cybertron": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "OpenPipe/mistral-ft-optimized-1227": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "perlthoughts/Falkor-7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "v1olet/v1olet_merged_dpo_7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Ba2han/BruinsV2-OpHermesNeu-11B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "DopeorNope/You_can_cry_Snowman-13B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "PistachioAlt/Synatra-MCS-7B-v0.3-RP-Slerp": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/MetaMath-una-cybertron-v2-bf16-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/OpenHermes-2.5-neural-chat-7b-v3-2-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "perlthoughts/Falkor-8x7B-MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "elinas/chronos007-70b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/MetaMath-NeuralHermes-2.5-Mistral-7B-Linear": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/MetaMath-neural-chat-7b-v3-2-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "diffnamehard/Mistral-CatMacaroni-slerp-uncensored-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/neural-chat-7b-v3-1-OpenHermes-2.5-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/MetaMath-NeuralHermes-2.5-Mistral-7B-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Walmart-the-bag/Misted-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "garage-bAInd/Camel-Platypus2-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "Weyaxi/OpenOrca-Zephyr-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "uukuguy/speechless-mistral-7b-dare-0.85": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
+    "DopeorNope/SOLARC-M-10.7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
+    "cloudyu/Mixtral_11Bx2_MoE_19B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
+    "DopeorNope/SOLARC-MOE-10.7Bx6 ": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
+    "DopeorNope/SOLARC-MOE-10.7Bx4": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
+    "gagan3012/MetaModelv2 ": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
+    "udkai/Turdus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "kodonho/Solar-OrcaDPO-Solar-Instruct-SLERP": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "kodonho/SolarM-SakuraSolar-SLERP": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "Yhyu13/LMCocktail-10.7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "mlabonne/NeuralMarcoro14-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "Neuronovo/neuronovo-7B-v0.2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "ryandt/MusingCaterpillar": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "Neuronovo/neuronovo-7B-v0.3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "SanjiWatsuki/Lelantos-DPO-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "bardsai/jaskier-7b-dpo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "cookinai/OpenCM-14": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "bardsai/jaskier-7b-dpo-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "jan-hq/supermario-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    # MoErges
+    "cloudyu/Yi-34Bx2-MoE-60B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "cloudyu/Mixtral_34Bx2_MoE_60B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "gagan3012/MetaModel_moe": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/SOLAR-math-2x10.7b-v0.2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "cloudyu/Mixtral_7Bx2_MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/SOLAR-math-2x10.7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/Orca-SOLAR-4x10.7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/piccolo-8x7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "cloudyu/Mixtral_7Bx4_MOE_24B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/laser-dolphin-mixtral-2x7b-dpo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    "macadeliccc/polyglot-math-4x7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
+    # Other - contamination mostly
+    "DopeorNope/COKAL-v1-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/566",
+    "CultriX/MistralTrix-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/556",
+    "Contamination/contaminated_proof_7b_v1.0": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/664",
+    "Contamination/contaminated_proof_7b_v1.0_safetensor": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/664",
+}
+
+# Models which have been requested by orgs to not be submitted on the leaderboard
+DO_NOT_SUBMIT_MODELS = [
+    "Voicelab/trurl-2-13b",  # trained on MMLU
+    "TigerResearch/tigerbot-70b-chat",  # per authors request
+    "TigerResearch/tigerbot-70b-chat-v2",  # per authors request
+    "TigerResearch/tigerbot-70b-chat-v4-4k",  # per authors request
+]
+
+
+def flag_models(leaderboard_data: list[dict]):
+    """Flags models based on external criteria or flagged status."""
+    for model_data in leaderboard_data:
+        # If a model is not flagged, use its "fullname" as a key
+        if model_data[AutoEvalColumn.not_flagged.name]:
+            flag_key = model_data[AutoEvalColumn.fullname.name]
+        else:
+             # Merges and moes are flagged 
+            flag_key = "merged"
+        
+        # Reverse the logic: Check for non-flagged models instead
+        if flag_key in FLAGGED_MODELS:
+            issue_num = FLAGGED_MODELS[flag_key].split("/")[-1]
+            issue_link = model_hyperlink(
+                FLAGGED_MODELS[flag_key],
+                f"See discussion #{issue_num}",
+            )
+            model_data[AutoEvalColumn.model.name] = (
+                f"{model_data[AutoEvalColumn.model.name]} has been flagged! {issue_link}"
+            )
+            model_data[AutoEvalColumn.not_flagged.name] = False
+        else:
+            model_data[AutoEvalColumn.not_flagged.name] = True
+
+
+def remove_forbidden_models(leaderboard_data: list[dict]):
+    """Removes models from the leaderboard based on the DO_NOT_SUBMIT list."""
+    indices_to_remove = []
+    for ix, model in enumerate(leaderboard_data):
+        if model[AutoEvalColumn.fullname.name] in DO_NOT_SUBMIT_MODELS:
+            indices_to_remove.append(ix)
+
+    # Remove the models from the list
+    for ix in reversed(indices_to_remove):
+        leaderboard_data.pop(ix)
+    return leaderboard_data
+
+
+def filter_models_flags(leaderboard_data: list[dict]):
+    leaderboard_data = remove_forbidden_models(leaderboard_data)
+    flag_models(leaderboard_data)
+

src/leaderboard/read_evals.py

@@ -0,0 +1,263 @@
+import json
+from pathlib import Path
+from json import JSONDecodeError
+import logging
+import math
+
+from dataclasses import dataclass, field
+from typing import Optional, Dict, List
+
+from tqdm import tqdm
+from tqdm.contrib.logging import logging_redirect_tqdm
+
+import numpy as np
+
+from src.display.formatting import make_clickable_model
+from src.display.utils import AutoEvalColumn, ModelType, Precision, Tasks, WeightType, parse_datetime
+
+# Configure logging
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+@dataclass
+class EvalResult:
+    # Also see src.display.utils.AutoEvalColumn for what will be displayed.
+    eval_name: str # org_model_precision (uid)
+    full_model: str # org/model (path on hub)
+    org: Optional[str]
+    model: str
+    revision: str # commit hash, "" if main
+    results: Dict[str, float]
+    precision: Precision = Precision.Unknown
+    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
+    weight_type: WeightType = WeightType.Original
+    architecture: str = "Unknown" # From config file
+    license: str = "?"
+    likes: int = 0
+    num_params: int = 0
+    date: str = "" # submission date of request file
+    still_on_hub: bool = True
+    is_merge: bool = False
+    not_flagged: bool = False
+    status: str = "FINISHED"
+    # List of tags, initialized to a new empty list for each instance to avoid the pitfalls of mutable default arguments.
+    tags: List[str] = field(default_factory=list)
+ 
+    
+    @classmethod
+    def init_from_json_file(cls, json_filepath: str) -> 'EvalResult':
+        with open(json_filepath, 'r') as fp:
+            data = json.load(fp)
+
+        config = data.get("config_general", {})
+        precision = Precision.from_str(config.get("model_dtype", "unknown"))
+        org_and_model = config.get("model_name", "").split("/", 1)
+        org = org_and_model[0] if len(org_and_model) > 1 else None
+        model = org_and_model[-1]
+        if len(org_and_model) == 1:
+            org = None
+            model = org_and_model[0]
+            result_key = f"{model}_{precision.value.name}"
+        else:
+            org = org_and_model[0]
+            model = org_and_model[1]
+            result_key = f"{org}_{model}_{precision.value.name}"
+        full_model = "/".join(org_and_model)
+
+        results = cls.extract_results(data)  # Properly call the method to extract results
+
+        return cls(
+            eval_name=result_key,
+            full_model=full_model,
+            org=org,
+            model=model,
+            results=results,
+            precision=precision,
+            revision=config.get("model_sha", "")
+        )
+
+    @staticmethod
+    def extract_results(data: Dict) -> Dict[str, float]:
+        """
+        Extract and process benchmark results from a given dict.
+
+        Parameters:
+        - data (Dict): A dictionary containing benchmark data. This dictionary must
+        include 'versions' and 'results' keys with respective sub-data.
+
+        Returns:
+        - Dict[str, float]: A dictionary where keys are benchmark names and values
+        are the processed average scores as percentages.
+
+        Notes:
+        - The method specifically checks for certain benchmark names to skip outdated entries.
+        - Handles NaN values by setting the corresponding benchmark result to 0.0.
+        - Averages scores across metrics for benchmarks found in the data, in a percentage format.
+        """
+        results = {}
+        for task in Tasks:
+            task = task.value
+            # We skip old mmlu entries
+            if task.benchmark == "hendrycksTest":
+                for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]:
+                    if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0:
+                        continue
+
+            # Some benchamrk values are NaNs, mostly truthfulQA
+            # Would be more optimal (without the whole dict itertion) if benchmark name was same as key in results
+            # e.g. not harness|truthfulqa:mc|0 but truthfulqa:mc
+            for k, v in data["results"].items():
+                if task.benchmark in k:
+                    if math.isnan(float(v[task.metric])):
+                        results[task.benchmark] = 0.0
+                        continue
+
+            # We average all scores of a given metric (mostly for mmlu)
+            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k])
+            if accs.size == 0 or any([acc is None for acc in accs]):
+                continue
+
+            mean_acc = np.mean(accs) * 100.0
+            results[task.benchmark] = mean_acc
+        
+        return results
+
+
+    def update_with_request_file(self, requests_path):
+        """Finds the relevant request file for the current model and updates info with it."""
+        try:
+            request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
+            if request_file is None:
+                logging.warning(f"No request file for {self.org}/{self.model}")
+                self.status = "FAILED"
+                return
+            
+            with open(request_file, "r") as f:
+                request = json.load(f)
+            
+            self.model_type = ModelType.from_str(request.get("model_type", "Unknown"))
+            self.weight_type = WeightType[request.get("weight_type", "Original")]
+            self.num_params = int(request.get("params", 0))  # Ensuring type safety
+            self.date = request.get("submitted_time", "")
+            self.architecture = request.get("architectures", "Unknown")
+            self.status = request.get("status", "FAILED")
+            
+        except FileNotFoundError:
+            self.status = "FAILED"
+            logging.error(f"Request file: {request_file} not found for {self.org}/{self.model}")
+        except JSONDecodeError:
+            self.status = "FAILED"
+            logging.error(f"Error decoding JSON from the request file for {self.org}/{self.model}")
+        except KeyError as e:
+            self.status = "FAILED"
+            logging.error(f"Key error {e} in processing request file for {self.org}/{self.model}")
+        except Exception as e:  # Catch-all for any other unexpected exceptions
+            self.status = "FAILED"
+            logging.error(f"Unexpected error {e} for {self.org}/{self.model}")
+
+
+    def update_with_dynamic_file_dict(self, file_dict):
+        """Update object attributes based on the provided dictionary, with error handling for missing keys and type validation."""
+        # Default values set for optional or potentially missing keys.
+        self.license = file_dict.get("license", "?")
+        self.likes = int(file_dict.get("likes", 0))  # Ensure likes is treated as an integer
+        self.still_on_hub = file_dict.get("still_on_hub", False)  # Default to False if key is missing
+        self.tags = file_dict.get("tags", [])
+        
+        # Calculate `flagged` only if 'tags' is not empty and avoid calculating each time
+        self.not_flagged = not (any("flagged" in tag for tag in self.tags))
+
+
+    def to_dict(self):
+        """Converts the Eval Result to a dict compatible with our dataframe display"""
+        average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
+        data_dict = {
+            "eval_name": self.eval_name,  # not a column, just a save name,
+            AutoEvalColumn.precision.name: self.precision.value.name,
+            AutoEvalColumn.model_type.name: self.model_type.value.name,
+            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
+            AutoEvalColumn.weight_type.name: self.weight_type.value.name,
+            AutoEvalColumn.architecture.name: self.architecture,
+            AutoEvalColumn.model.name: make_clickable_model(self.full_model),
+            AutoEvalColumn.fullname.name: self.full_model,
+            AutoEvalColumn.revision.name: self.revision,
+            AutoEvalColumn.average.name: average,
+            AutoEvalColumn.license.name: self.license,
+            AutoEvalColumn.likes.name: self.likes,
+            AutoEvalColumn.params.name: self.num_params,
+            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
+            AutoEvalColumn.merged.name: not( "merge" in self.tags if self.tags else False),
+            AutoEvalColumn.moe.name: not ( ("moe" in self.tags if self.tags else False) or "moe" in self.full_model.lower()) ,
+            AutoEvalColumn.not_flagged.name: self.not_flagged,
+        }
+
+        for task in Tasks:
+            data_dict[task.value.col_name] = self.results[task.value.benchmark]
+
+        return data_dict
+    
+
+def get_request_file_for_model(requests_path, model_name, precision):
+    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
+    requests_path = Path(requests_path)
+    pattern = f"{model_name}_eval_request_*.json"
+    
+    # Using pathlib to find files matching the pattern
+    request_files = list(requests_path.glob(pattern))
+    
+    # Sort the files by name in descending order to mimic 'reverse=True'
+    request_files.sort(reverse=True)
+
+    # Select the correct request file based on 'status' and 'precision'
+    request_file = None
+    for request_file in request_files:
+        with request_file.open("r") as f:
+            req_content = json.load(f)
+            if req_content["status"] == "FINISHED" and req_content["precision"] == precision.split(".")[-1]:
+                request_file = str(request_file)
+    
+    # Return empty string if no file found that matches criteria
+    return request_file
+
+
+def get_raw_eval_results(results_path: str, requests_path: str, dynamic_path: str) -> list[EvalResult]:
+    """From the path of the results folder root, extract all needed info for results"""
+    with open(dynamic_path) as f:
+        dynamic_data = json.load(f)
+    
+    results_path = Path(results_path)
+    model_files = list(results_path.rglob('results_*.json'))
+    model_files.sort(key=lambda file: parse_datetime(file.stem.removeprefix("results_")))
+
+    eval_results = {}
+    # Wrap model_files iteration with tqdm for progress display
+    for model_result_filepath in tqdm(model_files, desc="Processing model files"):
+        # Creation of result
+        eval_result = EvalResult.init_from_json_file(model_result_filepath)
+        with logging_redirect_tqdm():
+            eval_result.update_with_request_file(requests_path)
+
+        if eval_result.full_model in dynamic_data:
+            eval_result.update_with_dynamic_file_dict(dynamic_data[eval_result.full_model])
+            # Hardcoding because of gating problem
+            if any([org in eval_result.full_model for org in ["meta-llama/", "google/", "tiiuae/"]]):
+                eval_result.still_on_hub = True
+
+        # Store results of same eval together
+        eval_name = eval_result.eval_name
+        if eval_name in eval_results.keys():
+            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
+        else:
+            eval_results[eval_name] = eval_result
+
+    results = []
+    for k, v in eval_results.items():
+        try:
+            if v.status == "FINISHED":
+                v.to_dict()  # we test if the dict version is complete
+                results.append(v)
+        except KeyError as e:
+            logging.error(f"Error while checking model {k} {v.date} json, no key: {e}")  # not all eval values present
+            continue
+
+    return results
+

src/populate.py

@@ -0,0 +1,54 @@
+import json
+import os
+import pathlib
+import pandas as pd
+from src.display.formatting import has_no_nan_values, make_clickable_model
+from src.display.utils import AutoEvalColumn, EvalQueueColumn, baseline_row
+from src.leaderboard.filter_models import filter_models_flags
+from src.leaderboard.read_evals import get_raw_eval_results
+from src.display.utils import load_json_data
+
+
+def _process_model_data(entry, model_name_key="model", revision_key="revision"):
+    """Enrich model data with clickable links and revisions."""
+    entry[EvalQueueColumn.model.name] = make_clickable_model(entry.get(model_name_key, ""))
+    entry[EvalQueueColumn.revision.name] = entry.get(revision_key, "main")
+    return entry
+
+
+def get_evaluation_queue_df(save_path, cols):
+    """Generate dataframes for pending, running, and finished evaluation entries."""
+    save_path = pathlib.Path(save_path)
+    all_evals = []
+
+    for path in save_path.rglob('*.json'):
+        data = load_json_data(path)
+        if data:
+            all_evals.append(_process_model_data(data))
+
+    # Organizing data by status
+    status_map = {
+        "PENDING": ["PENDING", "RERUN"],
+        "RUNNING": ["RUNNING"],
+        "FINISHED": ["FINISHED", "PENDING_NEW_EVAL"],
+    }
+    status_dfs = {status: [] for status in status_map}
+    for eval_data in all_evals:
+        for status, extra_statuses in status_map.items():
+            if eval_data["status"] in extra_statuses:
+                status_dfs[status].append(eval_data)
+
+    return tuple(pd.DataFrame(status_dfs[status], columns=cols) for status in ["FINISHED", "RUNNING", "PENDING"])
+
+
+def get_leaderboard_df(results_path, requests_path, dynamic_path, cols, benchmark_cols):
+    """Retrieve and process leaderboard data."""
+    raw_data = get_raw_eval_results(results_path, requests_path, dynamic_path)
+    all_data_json = [model.to_dict() for model in raw_data] + [baseline_row]
+    filter_models_flags(all_data_json)
+
+    df = pd.DataFrame.from_records(all_data_json)
+    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
+    df = df[cols].round(decimals=2)
+    df = df[has_no_nan_values(df, benchmark_cols)]
+    return raw_data, df

src/scripts/create_request_file.py

@@ -0,0 +1,92 @@
+import json
+import os
+import pprint
+from datetime import datetime, timezone
+
+import click
+from colorama import Fore
+from huggingface_hub import HfApi, snapshot_download
+
+from src.display.utils import ModelType, WeightType
+from src.submission.check_validity import get_model_size
+
+EVAL_REQUESTS_PATH = "eval-queue"
+QUEUE_REPO = "open-llm-leaderboard/requests"
+
+precisions = ("float16", "bfloat16", "8bit (LLM.int8)", "4bit (QLoRA / FP4)", "GPTQ")
+model_types = [e.name for e in ModelType]
+weight_types = [e.name for e in WeightType]
+
+
+def main():
+    api = HfApi()
+    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+    snapshot_download(repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH, repo_type="dataset")
+
+    model_name = click.prompt("Enter model name")
+    revision = click.prompt("Enter revision", default="main")
+    precision = click.prompt("Enter precision", default="float16", type=click.Choice(precisions))
+    model_type = click.prompt("Enter model type", type=click.Choice(model_types))
+    weight_type = click.prompt("Enter weight type", default="Original", type=click.Choice(weight_types))
+    base_model = click.prompt("Enter base model", default="")
+    status = click.prompt("Enter status", default="FINISHED")
+
+    try:
+        model_info = api.model_info(repo_id=model_name, revision=revision)
+    except Exception as e:
+        print(f"{Fore.RED}Could not find model info for {model_name} on the Hub\n{e}{Fore.RESET}")
+        return 1
+
+    model_size = get_model_size(model_info=model_info, precision=precision)
+
+    try:
+        license = model_info.cardData["license"]
+    except Exception:
+        license = "?"
+
+    eval_entry = {
+        "model": model_name,
+        "base_model": base_model,
+        "revision": model_info.sha, # force to use the exact model commit 
+        "private": False,
+        "precision": precision,
+        "weight_type": weight_type,
+        "status": status,
+        "submitted_time": current_time,
+        "model_type": model_type,
+        "likes": model_info.likes,
+        "params": model_size,
+        "license": license,
+    }
+
+    user_name = ""
+    model_path = model_name
+    if "/" in model_name:
+        user_name = model_name.split("/")[0]
+        model_path = model_name.split("/")[1]
+
+    pprint.pprint(eval_entry)
+
+    if click.confirm("Do you want to continue? This request file will be pushed to the hub"):
+        click.echo("continuing...")
+
+        out_dir = f"{EVAL_REQUESTS_PATH}/{user_name}"
+        os.makedirs(out_dir, exist_ok=True)
+        out_path = f"{out_dir}/{model_path}_eval_request_{False}_{precision}_{weight_type}.json"
+
+        with open(out_path, "w") as f:
+            f.write(json.dumps(eval_entry))
+
+        api.upload_file(
+            path_or_fileobj=out_path,
+            path_in_repo=out_path.split(f"{EVAL_REQUESTS_PATH}/")[1],
+            repo_id=QUEUE_REPO,
+            repo_type="dataset",
+            commit_message=f"Add {model_name} to eval queue",
+        )
+    else:
+        click.echo("aborting...")
+
+
+if __name__ == "__main__":
+    main()

src/scripts/update_all_request_files.py

@@ -0,0 +1,99 @@
+import json
+import os
+import time
+
+from huggingface_hub import snapshot_download
+
+from src.envs import API, DYNAMIC_INFO_FILE_PATH, DYNAMIC_INFO_PATH, DYNAMIC_INFO_REPO, EVAL_REQUESTS_PATH, H4_TOKEN
+from src.submission.check_validity import check_model_card, get_model_tags, is_model_on_hub
+
+
+def update_one_model(model_id, data, models_on_the_hub):
+    # Model no longer on the hub at all
+    if model_id not in models_on_the_hub:
+        data["still_on_hub"] = False
+        data["likes"] = 0
+        data["downloads"] = 0
+        data["created_at"] = ""
+        data["tags"] = []
+        return data
+
+    # Grabbing model parameters
+    model_cfg = models_on_the_hub[model_id]
+    data["likes"] = model_cfg.likes
+    data["downloads"] = model_cfg.downloads
+    data["created_at"] = str(model_cfg.created_at)
+    data["license"] = model_cfg.card_data.license if model_cfg.card_data is not None else ""
+
+    # Grabbing model details
+    model_name = model_id
+    if model_cfg.card_data is not None and model_cfg.card_data.base_model is not None:
+        if isinstance(model_cfg.card_data.base_model, str):
+            model_name = model_cfg.card_data.base_model  # for adapters, we look at the parent model
+    still_on_hub, _, _ = is_model_on_hub(
+        model_name=model_name,
+        revision=data.get("revision"),
+        trust_remote_code=True,
+        test_tokenizer=False,
+        token=H4_TOKEN,
+    )
+    # If the model doesn't have a model card or a license, we consider it's deleted
+    if still_on_hub:
+        try:
+            status, _, model_card = check_model_card(model_id)
+            if status is False:
+                still_on_hub = False
+        except Exception:
+            model_card = None
+            still_on_hub = False
+    data["still_on_hub"] = still_on_hub
+
+    tags = get_model_tags(model_card, model_id) if still_on_hub else []
+
+    data["tags"] = tags
+    return data
+
+
+def update_models(file_path, models_on_the_hub):
+    """
+    Search through all JSON files in the specified root folder and its subfolders,
+    and update the likes key in JSON dict from value of input dict
+    """
+    seen_models = []
+    with open(file_path, "r") as f:
+        model_infos = json.load(f)
+        for model_id in model_infos.keys():
+            seen_models.append(model_id)
+            model_infos[model_id] = update_one_model(
+                model_id=model_id, data=model_infos[model_id], models_on_the_hub=models_on_the_hub
+            )
+
+    # If new requests files have been created since we started all this
+    # we grab them
+    all_models = []
+    try:
+        for ix, (root, _, files) in enumerate(os.walk(EVAL_REQUESTS_PATH)):
+            if ix == 0:
+                continue
+            for file in files:
+                if "eval_request" in file:
+                    path = root.split("/")[-1] + "/" + file.split("_eval_request")[0]
+                    all_models.append(path)
+    except Exception as e:
+        print(e)
+        pass
+
+    for model_id in all_models:
+        if model_id not in seen_models:
+            model_infos[model_id] = update_one_model(model_id=model_id, data={}, models_on_the_hub=models_on_the_hub)
+
+    with open(file_path, "w") as f:
+        json.dump(model_infos, f, indent=2)
+
+
+def update_dynamic_files():
+    # from gen import gen_answer,gen_judgment\
+    import subprocess
+    subprocess.Popen('python3 ../gen/gen_judgement.py')
+
+    subprocess.Popen('python3 ../gen/show_result.py --output')

src/submission/check_validity.py

@@ -0,0 +1,178 @@
+import json
+import os
+import re
+from collections import defaultdict
+from datetime import datetime, timedelta, timezone
+
+import huggingface_hub
+from huggingface_hub import ModelCard
+from huggingface_hub.hf_api import ModelInfo, get_safetensors_metadata
+from transformers import AutoConfig, AutoTokenizer
+
+from src.envs import HAS_HIGHER_RATE_LIMIT
+
+
+# ht to @Wauplin, thank you for the snippet!
+# See https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/317
+def check_model_card(repo_id: str) -> tuple[bool, str]:
+    # Returns operation status, and error message
+    try:
+        card = ModelCard.load(repo_id)
+    except huggingface_hub.utils.EntryNotFoundError:
+        return False, "Please add a model card to your model to explain how you trained/fine-tuned it.", None
+
+    # Enforce license metadata
+    if card.data.license is None:
+        if not ("license_name" in card.data and "license_link" in card.data):
+            return (
+                False,
+                (
+                    "License not found. Please add a license to your model card using the `license` metadata or a"
+                    " `license_name`/`license_link` pair."
+                ),
+                None,
+            )
+
+    # Enforce card content
+    if len(card.text) < 200:
+        return False, "Please add a description to your model card, it is too short.", None
+
+    return True, "", card
+
+
+def is_model_on_hub(
+    model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False
+) -> tuple[bool, str, AutoConfig]:
+    try:
+        config = AutoConfig.from_pretrained(
+            model_name, revision=revision, trust_remote_code=trust_remote_code, token=token
+        )  # , force_download=True)
+        if test_tokenizer:
+            try:
+                tk = AutoTokenizer.from_pretrained(
+                    model_name, revision=revision, trust_remote_code=trust_remote_code, token=token
+                )
+            except ValueError as e:
+                return (False, f"uses a tokenizer which is not in a transformers release: {e}", None)
+            except Exception:
+                return (
+                    False,
+                    "'s tokenizer cannot be loaded. Is your tokenizer class in a stable transformers release, and correctly configured?",
+                    None,
+                )
+        return True, None, config
+
+    except ValueError:
+        return (
+            False,
+            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
+            None,
+        )
+
+    except Exception as e:
+        if "You are trying to access a gated repo." in str(e):
+            return True, "uses a gated model.", None
+        return False, f"was not found or misconfigured on the hub! Error raised was {e.args[0]}", None
+
+
+def get_model_size(model_info: ModelInfo, precision: str):
+    size_pattern = re.compile(r"(\d+\.)?\d+(b|m)")
+    safetensors = None
+    try:
+        safetensors = get_safetensors_metadata(model_info.id)
+    except Exception as e:
+        print(e)
+
+    if safetensors is not None:
+        model_size = round(sum(safetensors.parameter_count.values()) / 1e9, 3)
+    else:
+        try:
+            size_match = re.search(size_pattern, model_info.id.lower())
+            model_size = size_match.group(0)
+            model_size = round(float(model_size[:-1]) if model_size[-1] == "b" else float(model_size[:-1]) / 1e3, 3)
+        except AttributeError:
+            return 0  # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py
+
+    size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.id.lower()) else 1
+    model_size = size_factor * model_size
+    return model_size
+
+
+def get_model_arch(model_info: ModelInfo):
+    return model_info.config.get("architectures", "Unknown")
+
+
+def user_submission_permission(org_or_user, users_to_submission_dates, rate_limit_period, rate_limit_quota):
+    if org_or_user not in users_to_submission_dates:
+        return True, ""
+    submission_dates = sorted(users_to_submission_dates[org_or_user])
+
+    time_limit = (datetime.now(timezone.utc) - timedelta(days=rate_limit_period)).strftime("%Y-%m-%dT%H:%M:%SZ")
+    submissions_after_timelimit = [d for d in submission_dates if d > time_limit]
+
+    num_models_submitted_in_period = len(submissions_after_timelimit)
+    if org_or_user in HAS_HIGHER_RATE_LIMIT:
+        rate_limit_quota = 2 * rate_limit_quota
+
+    if num_models_submitted_in_period > rate_limit_quota:
+        error_msg = f"Organisation or user `{org_or_user}`"
+        error_msg += f"already has {num_models_submitted_in_period} model requests submitted to the leaderboard "
+        error_msg += f"in the last {rate_limit_period} days.\n"
+        error_msg += (
+            "Please wait a couple of days before resubmitting, so that everybody can enjoy using the leaderboard 🤗"
+        )
+        return False, error_msg
+    return True, ""
+
+
+def already_submitted_models(requested_models_dir: str) -> set[str]:
+    depth = 1
+    file_names = []
+    users_to_submission_dates = defaultdict(list)
+
+    for root, _, files in os.walk(requested_models_dir):
+        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
+        if current_depth == depth:
+            for file in files:
+                if not file.endswith(".json"):
+                    continue
+                with open(os.path.join(root, file), "r") as f:
+                    info = json.load(f)
+                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
+
+                    # Select organisation
+                    if info["model"].count("/") == 0 or "submitted_time" not in info:
+                        continue
+                    organisation, _ = info["model"].split("/")
+                    users_to_submission_dates[organisation].append(info["submitted_time"])
+
+    return set(file_names), users_to_submission_dates
+
+
+def get_model_tags(model_card, model: str):
+    is_merge_from_metadata = False
+    is_moe_from_metadata = False
+
+    tags = []
+    if model_card is None:
+        return tags
+    if model_card.data.tags:
+        is_merge_from_metadata = any(
+            [tag in model_card.data.tags for tag in ["merge", "moerge", "mergekit", "lazymergekit"]]
+        )
+        is_moe_from_metadata = any([tag in model_card.data.tags for tag in ["moe", "moerge"]])
+
+    is_merge_from_model_card = any(
+        keyword in model_card.text.lower() for keyword in ["merged model", "merge model", "moerge"]
+    )
+    if is_merge_from_model_card or is_merge_from_metadata:
+        tags.append("merge")
+    is_moe_from_model_card = any(keyword in model_card.text.lower() for keyword in ["moe", "mixtral"])
+    # Hardcoding because of gating problem
+    if "Qwen/Qwen1.5-32B" in model:
+        is_moe_from_model_card = False
+    is_moe_from_name = "moe" in model.lower().replace("/", "-").replace("_", "-").split("-")
+    if is_moe_from_model_card or is_moe_from_name or is_moe_from_metadata:
+        tags.append("moe")
+
+    return tags

src/submission/submit.py

@@ -0,0 +1,191 @@
+import json
+import os
+from datetime import datetime, timezone
+
+from huggingface_hub import snapshot_download
+
+from src.display.formatting import styled_error, styled_message, styled_warning
+from src.envs import (
+    API,
+    DYNAMIC_INFO_FILE_PATH,
+    DYNAMIC_INFO_PATH,
+    DYNAMIC_INFO_REPO,
+    EVAL_REQUESTS_PATH,
+    H4_TOKEN,
+    QUEUE_REPO,
+    RATE_LIMIT_PERIOD,
+    RATE_LIMIT_QUOTA,
+)
+# from src.leaderboard.filter_models import DO_NOT_SUBMIT_MODELS
+# from src.submission.check_validity import (
+#     already_submitted_models,
+#     check_model_card,
+#     get_model_size,
+#     get_model_tags,
+#     is_model_on_hub,
+#     user_submission_permission,
+# )
+
+REQUESTED_MODELS = None
+USERS_TO_SUBMISSION_DATES = None
+
+
+def add_new_eval(
+    model: str,
+):
+    # global REQUESTED_MODELS
+    # global USERS_TO_SUBMISSION_DATES
+    # if not REQUESTED_MODELS:
+    #     REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)
+
+
+    # user_name = ""
+    # model_path = model
+    # if "/" in model:
+    #     user_name = model.split("/")[0]
+    #     model_path = model.split("/")[1]
+
+    # # precision = precision.split(" ")[0]
+    # current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+    # if model_type is None or model_type == "":
+    #     return styled_error("Please select a model type.")
+
+    # # Is the user rate limited?
+    # if user_name != "":
+    #     user_can_submit, error_msg = user_submission_permission(
+    #         user_name, USERS_TO_SUBMISSION_DATES, RATE_LIMIT_PERIOD, RATE_LIMIT_QUOTA
+    #     )
+    #     if not user_can_submit:
+    #         return styled_error(error_msg)
+
+    # Did the model authors forbid its submission to the leaderboard?
+    # if model in DO_NOT_SUBMIT_MODELS or base_model in DO_NOT_SUBMIT_MODELS:
+    #     return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.")
+
+    # if model == "CohereForAI/c4ai-command-r-plus":
+    #     return styled_warning(
+    #         "This model cannot be submitted manually on the leaderboard before the transformers release."
+    #     )
+
+    # # Does the model actually exist?
+    # if revision == "":
+    #     revision = "main"
+
+    # # Is the model on the hub?
+    # if weight_type in ["Delta", "Adapter"]:
+    #     base_model_on_hub, error, _ = is_model_on_hub(
+    #         model_name=base_model, revision=revision, token=H4_TOKEN, test_tokenizer=True
+    #     )
+    #     if not base_model_on_hub:
+    #         return styled_error(f'Base model "{base_model}" {error}')
+
+    # architecture = "?"
+    # downloads = 0
+    # created_at = ""
+    # if not weight_type == "Adapter":
+    #     model_on_hub, error, model_config = is_model_on_hub(model_name=model, revision=revision, test_tokenizer=True)
+    #     if not model_on_hub or model_config is None:
+    #         return styled_error(f'Model "{model}" {error}')
+    #     if model_config is not None:
+    #         architectures = getattr(model_config, "architectures", None)
+    #         if architectures:
+    #             architecture = ";".join(architectures)
+    #         downloads = getattr(model_config, "downloads", 0)
+    #         created_at = getattr(model_config, "created_at", "")
+
+    # Is the model info correctly filled?
+    # try:
+    #     model_info = API.model_info(repo_id=model, revision=revision)
+    # except Exception:
+    #     return styled_error("Could not get your model information. Please fill it up properly.")
+
+    # model_size = get_model_size(model_info=model_info, precision=precision)
+
+    # Were the model card and license filled?
+    # try:
+    #     license = model_info.cardData["license"]
+    # except Exception:
+    #     return styled_error("Please select a license for your model")
+
+    # modelcard_OK, error_msg, model_card = check_model_card(model)
+    # if not modelcard_OK:
+    #     return styled_error(error_msg)
+
+    # tags = get_model_tags(model_card, model)
+
+    # # Seems good, creating the eval
+    # print("Adding new eval")
+
+    # eval_entry = {
+    #     "model": model,
+    #     # "base_model": base_model,
+    #     # "revision": model_info.sha, # force to use the exact model commit
+    #     # "private": private,
+    #     # "precision": precision,
+    #     # "params": model_size,
+    #     # "architectures": architecture,
+    #     # "weight_type": weight_type,
+    #     "status": "PENDING",
+    #     # "submitted_time": current_time,
+    #     # "model_type": model_type,
+    #     "job_id": -1,
+    #     "job_start_time": None,
+    # }
+
+    # supplementary_info = {
+    #     "likes": model_info.likes,
+    #     "license": license,
+    #     "still_on_hub": True,
+    #     "tags": tags,
+    #     "downloads": downloads,
+    #     "created_at": created_at,
+    # }
+
+    # # Check for duplicate submission
+    # if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
+    #     return styled_warning("This model has been already submitted.")
+
+    # print("Creating eval file")
+    # OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
+    # os.makedirs(OUT_DIR, exist_ok=True)
+    # out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json"
+
+    # with open(out_path, "w") as f:
+    #     f.write(json.dumps(eval_entry))
+
+    # print("Uploading eval file")
+    # API.upload_file(
+    #     path_or_fileobj=out_path,
+    #     path_in_repo=out_path.split("eval-queue/")[1],
+    #     repo_id=QUEUE_REPO,
+    #     repo_type="dataset",
+    #     commit_message=f"Add {model} to eval queue",
+    # )
+
+    # We want to grab the latest version of the submission file to not accidentally overwrite it
+    # snapshot_download(
+    #     repo_id=DYNAMIC_INFO_REPO, local_dir=DYNAMIC_INFO_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
+    # )
+
+    # with open(DYNAMIC_INFO_FILE_PATH) as f:
+    #     all_supplementary_info = json.load(f)
+
+    # # all_supplementary_info[model] = supplementary_info
+    # with open(DYNAMIC_INFO_FILE_PATH, "w") as f:
+    #     json.dump(all_supplementary_info, f, indent=2)
+
+    # API.upload_file(
+    #     path_or_fileobj=DYNAMIC_INFO_FILE_PATH,
+    #     path_in_repo=DYNAMIC_INFO_FILE_PATH.split("/")[-1],
+    #     repo_id=DYNAMIC_INFO_REPO,
+    #     repo_type="dataset",
+    #     commit_message=f"Add {model} to dynamic info queue",
+    # )
+
+    # # Remove the local file
+    # os.remove(out_path)
+
+    return styled_message(
+        "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour."
+    )

src/tools/collections.py

@@ -0,0 +1,76 @@
+import pandas as pd
+from huggingface_hub import add_collection_item, delete_collection_item, get_collection, update_collection_item
+from huggingface_hub.utils._errors import HfHubHTTPError
+from pandas import DataFrame
+
+from src.display.utils import AutoEvalColumn, ModelType
+from src.envs import H4_TOKEN, PATH_TO_COLLECTION
+
+# Specific intervals for the collections
+intervals = {
+    "1B": pd.Interval(0, 1.5, closed="right"),
+    "3B": pd.Interval(2.5, 3.5, closed="neither"),
+    "7B": pd.Interval(6, 8, closed="neither"),
+    "13B": pd.Interval(10, 14, closed="neither"),
+    "30B": pd.Interval(25, 35, closed="neither"),
+    "65B": pd.Interval(60, 70, closed="neither"),
+}
+
+
+def _filter_by_type_and_size(df, model_type, size_interval):
+    """Filter DataFrame by model type and parameter size interval."""
+    type_emoji = model_type.value.symbol[0]
+    filtered_df = df[df[AutoEvalColumn.model_type_symbol.name] == type_emoji]
+    params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce")
+    mask = params_column.apply(lambda x: x in size_interval)
+    return filtered_df.loc[mask]
+
+
+def _add_models_to_collection(collection, models, model_type, size):
+    """Add best models to the collection and update positions."""
+    cur_len_collection = len(collection.items)
+    for ix, model in enumerate(models, start=1):
+        try:
+            collection = add_collection_item(
+                PATH_TO_COLLECTION,
+                item_id=model,
+                item_type="model",
+                exists_ok=True,
+                note=f"Best {model_type.to_str(' ')} model of around {size} on the leaderboard today!",
+                token=H4_TOKEN,
+            )
+            # Ensure position is correct if item was added
+            if len(collection.items) > cur_len_collection:
+                item_object_id = collection.items[-1].item_object_id
+                update_collection_item(collection_slug=PATH_TO_COLLECTION, item_object_id=item_object_id, position=ix)
+                cur_len_collection = len(collection.items)
+            break  # assuming we only add the top model
+        except HfHubHTTPError:
+            continue
+
+
+def update_collections(df: DataFrame):
+    """Update collections by filtering and adding the best models."""
+    collection = get_collection(collection_slug=PATH_TO_COLLECTION, token=H4_TOKEN)
+    cur_best_models = []
+
+    for model_type in ModelType:
+        if not model_type.value.name:
+            continue
+        for size, interval in intervals.items():
+            filtered_df = _filter_by_type_and_size(df, model_type, interval)
+            best_models = list(
+                filtered_df.sort_values(AutoEvalColumn.average.name, ascending=False)[AutoEvalColumn.fullname.name][:10]
+            )
+            print(model_type.value.symbol, size, best_models)
+            _add_models_to_collection(collection, best_models, model_type, size)
+            cur_best_models.extend(best_models)
+
+    # Cleanup
+    existing_models = {item.item_id for item in collection.items}
+    to_remove = existing_models - set(cur_best_models)
+    for item_id in to_remove:
+        try:
+            delete_collection_item(collection_slug=PATH_TO_COLLECTION, item_object_id=item_id, token=H4_TOKEN)
+        except HfHubHTTPError:
+            continue

src/tools/model_backlinks.py

@@ -0,0 +1,1309 @@
+models = [
+    "uni-tianyan/Uni-TianYan",
+    "fangloveskari/ORCA_LLaMA_70B_QLoRA",
+    "garage-bAInd/Platypus2-70B-instruct",
+    "upstage/Llama-2-70b-instruct-v2",
+    "fangloveskari/Platypus_QLoRA_LLaMA_70b",
+    "yeontaek/llama-2-70B-ensemble-v5",
+    "TheBloke/Genz-70b-GPTQ",
+    "TheBloke/Platypus2-70B-Instruct-GPTQ",
+    "psmathur/model_007",
+    "yeontaek/llama-2-70B-ensemble-v4",
+    "psmathur/orca_mini_v3_70b",
+    "ehartford/Samantha-1.11-70b",
+    "MayaPH/GodziLLa2-70B",
+    "psmathur/model_007_v2",
+    "chargoddard/MelangeA-70b",
+    "ehartford/Samantha-1.1-70b",
+    "psmathur/model_009",
+    "upstage/Llama-2-70b-instruct",
+    "yeontaek/llama-2-70B-ensemble-v7",
+    "yeontaek/llama-2-70B-ensemble-v6",
+    "chargoddard/MelangeB-70b",
+    "yeontaek/llama-2-70B-ensemble-v3",
+    "chargoddard/MelangeC-70b",
+    "garage-bAInd/Camel-Platypus2-70B",
+    "yeontaek/llama-2-70B-ensemble-v2",
+    "garage-bAInd/Camel-Platypus2-70B",
+    "migtissera/Synthia-70B-v1.2",
+    "v2ray/LLaMA-2-Wizard-70B-QLoRA",
+    "quantumaikr/llama-2-70b-fb16-orca-chat-10k",
+    "v2ray/LLaMA-2-Wizard-70B-QLoRA",
+    "stabilityai/StableBeluga2",
+    "quantumaikr/llama-2-70b-fb16-guanaco-1k",
+    "garage-bAInd/Camel-Platypus2-70B",
+    "migtissera/Synthia-70B-v1.1",
+    "migtissera/Synthia-70B",
+    "psmathur/model_101",
+    "augtoma/qCammel70",
+    "augtoma/qCammel-70",
+    "augtoma/qCammel-70v1",
+    "augtoma/qCammel-70x",
+    "augtoma/qCammel-70-x",
+    "jondurbin/airoboros-l2-70b-gpt4-1.4.1",
+    "dfurman/llama-2-70b-dolphin-peft",
+    "jondurbin/airoboros-l2-70b-2.1",
+    "TheBloke/llama-2-70b-Guanaco-QLoRA-fp16",
+    "quantumaikr/QuantumLM-llama2-70B-Korean-LoRA",
+    "quantumaikr/quantumairk-llama-2-70B-instruct",
+    "psmathur/model_420",
+    "psmathur/model_51",
+    "garage-bAInd/Camel-Platypus2-70B",
+    "TheBloke/Airoboros-L2-70B-2.1-GPTQ",
+    "OpenAssistant/llama2-70b-oasst-sft-v10",
+    "garage-bAInd/Platypus2-70B",
+    "liuxiang886/llama2-70B-qlora-gpt4",
+    "upstage/llama-65b-instruct",
+    "quantumaikr/llama-2-70b-fb16-korean",
+    "NousResearch/Nous-Hermes-Llama2-70b",
+    "v2ray/LLaMA-2-Jannie-70B-QLoRA",
+    "jondurbin/airoboros-l2-70b-gpt4-m2.0",
+    "jondurbin/airoboros-l2-70b-gpt4-m2.0",
+    "OpenAssistant/llama2-70b-oasst-sft-v10",
+    "yeontaek/llama-2-70B-ensemble-v8",
+    "jondurbin/airoboros-l2-70b-gpt4-2.0",
+    "jarradh/llama2_70b_chat_uncensored",
+    "WizardLM/WizardMath-70B-V1.0",
+    "jordiclive/Llama-2-70b-oasst-1-200",
+    "WizardLM/WizardMath-70B-V1.0",
+    "jondurbin/airoboros-l2-70b-gpt4-2.0",
+    "OpenLemur/lemur-70b-chat-v1",
+    "tiiuae/falcon-180B",
+    "tiiuae/falcon-180B",
+    "stabilityai/StableBeluga1-Delta",
+    "psmathur/model_42_70b",
+    "psmathur/test_42_70b",
+    "TheBloke/fiction.live-Kimiko-V2-70B-fp16",
+    "tiiuae/falcon-180B",
+    "WizardLM/WizardMath-70B-V1.0",
+    "tiiuae/falcon-180B-chat",
+    "jondurbin/airoboros-l2-70b-gpt4-2.0",
+    "ehartford/samantha-1.1-llama-33b",
+    "ajibawa-2023/scarlett-33b",
+    "ddobokki/Llama-2-70b-orca-200k",
+    "TheBloke/gpt4-alpaca-lora_mlp-65B-HF",
+    "tiiuae/falcon-180B-chat",
+    "tiiuae/falcon-180B-chat",
+    "tiiuae/falcon-180B",
+    "TheBloke/Lemur-70B-Chat-v1-GPTQ",
+    "NousResearch/Nous-Puffin-70B",
+    "WizardLM/WizardLM-70B-V1.0",
+    "WizardLM/WizardMath-70B-V1.0",
+    "meta-llama/Llama-2-70b-hf",
+    "TheBloke/Llama-2-70B-fp16",
+    "Weyaxi/llama-2-alpacagpt4-1000step",
+    "WizardLM/WizardLM-70B-V1.0",
+    "simsim314/WizardLM-70B-V1.0-HF",
+    "simsim314/WizardLM-70B-V1.0-HF",
+    "WizardLM/WizardLM-70B-V1.0",
+    "openbmb/UltraLM-65b",
+    "psmathur/model_420_preview",
+    "WizardLM/WizardLM-70B-V1.0",
+    "simsim314/WizardLM-70B-V1.0-HF",
+    "OpenBuddy/openbuddy-llama2-70b-v10.1-bf16",
+    "upstage/llama-30b-instruct-2048",
+    "jondurbin/airoboros-65b-gpt4-1.2",
+    "TheBloke/guanaco-65B-HF",
+    "jondurbin/airoboros-65b-gpt4-1.3",
+    "meta-llama/Llama-2-70b-chat-hf",
+    "ValiantLabs/ShiningValiant",
+    "Faradaylab/Aria-70B",
+    "lilloukas/GPlatty-30B",
+    "TheBloke/VicUnlocked-alpaca-65B-QLoRA-fp16",
+    "jondurbin/airoboros-65b-gpt4-1.4-peft",
+    "jondurbin/airoboros-65b-gpt4-1.4",
+    "jondurbin/airoboros-65b-gpt4-2.0",
+    "TheBloke/WizardLM-70B-V1.0-GPTQ",
+    "TheBloke/WizardLM-70B-V1.0-GPTQ",
+    "ariellee/SuperPlatty-30B",
+    "jondurbin/airoboros-65b-gpt4-1.4",
+    "jondurbin/airoboros-65b-gpt4-2.0",
+    "yeontaek/llama-2-70b-IA3-guanaco",
+    "CalderaAI/30B-Lazarus",
+    "Aspik101/trurl-2-13b-pl-instruct_unload",
+    "ehartford/WizardLM-33B-V1.0-Uncensored",
+    "ehartford/WizardLM-33B-V1.0-Uncensored",
+    "OpenBuddy/openbuddy-llama-65b-v8-bf16",
+    "Aspik101/llama-30b-instruct-2048-PL-lora",
+    "h2oai/h2ogpt-research-oasst1-llama-65b",
+    "Aspik101/llama-30b-instruct-2048-PL-lora",
+    "CalderaAI/30B-Epsilon",
+    "Aspik101/llama-30b-2048-instruct-PL-lora_unload",
+    "jondurbin/airoboros-65b-gpt4-m2.0",
+    "jondurbin/airoboros-65b-gpt4-m2.0",
+    "Aeala/Alpaca-elina-65b",
+    "TheBloke/robin-65b-v2-fp16",
+    "TheBloke/gpt4-alpaca-lora-30b-HF",
+    "TheBloke/Llama-2-70B-chat-GPTQ",
+    "upstage/llama-30b-instruct",
+    "OpenLemur/lemur-70b-v1",
+    "lmsys/vicuna-33b-v1.3",
+    "ausboss/llama-30b-supercot",
+    "ai-business/Luban-13B",
+    "Henk717/airochronos-33B",
+    "lmsys/vicuna-33b-v1.3",
+    "Henk717/airochronos-33B",
+    "bavest/fin-llama-33b-merged",
+    "jondurbin/airoboros-33b-gpt4-1.4",
+    "YeungNLP/firefly-llama-30b",
+    "Aspik101/30B-Lazarus-instruct-PL-lora_unload",
+    "uukuguy/speechless-llama2-luban-orca-platypus-13b",
+    "xxyyy123/test_merge_p_ov1_w0.66_w0.5_n1",
+    "jondurbin/airoboros-33b-gpt4-1.2",
+    "TheBloke/alpaca-lora-65B-HF",
+    "bofenghuang/vigogne-33b-instruct",
+    "yeontaek/llama-2-13B-ensemble-v5",
+    "garage-bAInd/Platypus-30B",
+    "Open-Orca/OpenOrca-Platypus2-13B",
+    "kajdun/viwaai-30b_v4",
+    "lilloukas/Platypus-30B",
+    "Open-Orca/OpenOrca-Platypus2-13B",
+    "Henk717/chronoboros-33B",
+    "jondurbin/airoboros-33b-2.1",
+    "HiTZ/alpaca-lora-65b-en-pt-es-ca",
+    "quantumaikr/QuantumLM-70B-hf",
+    "uukuguy/speechless-llama2-13b",
+    "uukuguy/speechless-llama2-hermes-orca-platypus-13b",
+    "openaccess-ai-collective/manticore-30b-chat-pyg-alpha",
+    "LLMs/WizardLM-30B-V1.0",
+    "TheBloke/WizardLM-30B-fp16",
+    "openaccess-ai-collective/hippogriff-30b-chat",
+    "concedo/Vicuzard-30B-Uncensored",
+    "TFLai/OpenOrca-Platypus2-13B-QLoRA-0.80-epoch",
+    "huggingface/llama-65b",
+    "huggyllama/llama-65b",
+    "gaodrew/gaodrew-llama-30b-instruct-2048-Open-Platypus-100steps",
+    "uukuguy/speechless-llama2-hermes-orca-platypus-wizardlm-13b",
+    "Sao10K/Mythical-Destroyer-V2-L2-13B",
+    "camel-ai/CAMEL-33B-Combined-Data",
+    "dsvv-cair/alpaca-cleaned-llama-30b-bf16",
+    "MetaIX/GPT4-X-Alpasta-30b",
+    "garage-bAInd/Stable-Platypus2-13B",
+    "TFLai/Luban-Platypus2-13B-QLora-0.80-epoch",
+    "TheBloke/OpenOrca-Platypus2-13B-GPTQ",
+    "IkariDev/Athena-tmp",
+    "OpenBuddyEA/openbuddy-llama-30b-v7.1-bf16",
+    "OpenBuddyEA/openbuddy-llama-30b-v7.1-bf16",
+    "Open-Orca/OpenOrcaxOpenChat-Preview2-13B",
+    "psmathur/model_007_13b_v2",
+    "Aspik101/Vicuzard-30B-Uncensored-instruct-PL-lora_unload",
+    "jondurbin/airoboros-33b-gpt4-m2.0",
+    "Sao10K/Mythical-Destroyer-L2-13B",
+    "TheBloke/Wizard-Vicuna-30B-Uncensored-fp16",
+    "ehartford/Wizard-Vicuna-30B-Uncensored",
+    "TFLai/Nova-13B",
+    "TheBloke/robin-33B-v2-fp16",
+    "totally-not-an-llm/PuddleJumper-13b",
+    "Aeala/VicUnlocked-alpaca-30b",
+    "Yhyu13/oasst-rlhf-2-llama-30b-7k-steps-hf",
+    "jondurbin/airoboros-33b-gpt4",
+    "jondurbin/airoboros-33b-gpt4-m2.0",
+    "tiiuae/falcon-40b-instruct",
+    "psmathur/orca_mini_v3_13b",
+    "Aeala/GPT4-x-AlpacaDente-30b",
+    "MayaPH/GodziLLa-30B",
+    "jondurbin/airoboros-33b-gpt4-m2.0",
+    "TFLai/SpeechlessV1-Nova-13B",
+    "yeontaek/llama-2-13B-ensemble-v4",
+    "ajibawa-2023/carl-33b",
+    "jondurbin/airoboros-33b-gpt4-2.0",
+    "TFLai/Stable-Platypus2-13B-QLoRA-0.80-epoch",
+    "jondurbin/airoboros-33b-gpt4-1.3",
+    "TehVenom/oasst-sft-6-llama-33b-xor-MERGED-16bit",
+    "TFLai/OrcaMini-Platypus2-13B-QLoRA-0.80-epoch",
+    "jondurbin/airoboros-33b-gpt4-2.0",
+    "chargoddard/Chronorctypus-Limarobormes-13b",
+    "jondurbin/airoboros-33b-gpt4-1.3",
+    "Open-Orca/OpenOrca-Platypus2-13B",
+    "FelixChao/vicuna-33b-coder",
+    "FelixChao/vicuna-33b-coder",
+    "Gryphe/MythoMix-L2-13b",
+    "Aeala/Enterredaas-33b",
+    "yeontaek/llama-2-13B-ensemble-v1",
+    "TFLai/OpenOrcaPlatypus2-Platypus2-13B-QLora-0.80-epoch",
+    "TFLai/Ensemble5-Platypus2-13B-QLora-0.80-epoch",
+    "yeontaek/llama-2-13B-ensemble-v3",
+    "TFLai/MythoMix-Platypus2-13B-QLoRA-0.80-epoch",
+    "yihan6324/llama2-13b-instructmining-40k-sharegpt",
+    "timdettmers/guanaco-33b-merged",
+    "TFLai/EnsembleV5-Nova-13B",
+    "circulus/Llama-2-13b-orca-v1",
+    "Undi95/ReMM-SLERP-L2-13B",
+    "Gryphe/MythoMax-L2-13b",
+    "stabilityai/StableBeluga-13B",
+    "circulus/Llama-2-13b-orca-v1",
+    "ehartford/WizardLM-30B-Uncensored",
+    "The-Face-Of-Goonery/huginnv1.2",
+    "TheBloke/OpenOrcaxOpenChat-Preview2-13B-GPTQ",
+    "Sao10K/Stheno-L2-13B",
+    "bofenghuang/vigogne-2-13b-instruct",
+    "The-Face-Of-Goonery/Huginn-13b-FP16",
+    "grimpep/L2-MythoMax22b-instruct-Falseblock",
+    "TFLai/Nous-Hermes-Platypus2-13B-QLoRA-0.80-epoch",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3-v4",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3-ensemble",
+    "Open-Orca/LlongOrca-13B-16k",
+    "Sao10K/Stheno-Inverted-L2-13B",
+    "garage-bAInd/Camel-Platypus2-13B",
+    "digitous/Alpacino30b",
+    "NousResearch/Nous-Hermes-Llama2-13b",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3-v3",
+    "TFLai/MythicalDestroyerV2-Platypus2-13B-QLora-0.80-epoch",
+    "TheBloke/VicUnlocked-30B-LoRA-HF",
+    "Undi95/Nous-Hermes-13B-Code",
+    "The-Face-Of-Goonery/Chronos-Beluga-v2-13bfp16",
+    "NousResearch/Nous-Hermes-Llama2-13b",
+    "Monero/WizardLM-Uncensored-SuperCOT-StoryTelling-30b",
+    "TheBloke/Wizard-Vicuna-30B-Uncensored-GPTQ",
+    "Open-Orca/OpenOrcaxOpenChat-Preview2-13B",
+    "Austism/chronos-hermes-13b-v2",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3-v2.1",
+    "yeontaek/Platypus2xOpenOrca-13B-IA3-v2",
+    "Gryphe/MythoLogic-L2-13b",
+    "augtoma/qCammel-13",
+    "YeungNLP/firefly-llama2-13b-v1.2",
+    "Aspik101/StableBeluga-13B-instruct-PL-lora_unload",
+    "andreaskoepf/llama2-13b-megacode2_min100",
+    "rombodawg/LosslessMegaCoder-llama2-13b-mini",
+    "yulan-team/YuLan-Chat-2-13b-fp16",
+    "elinas/chronos-33b",
+    "YeungNLP/firefly-llama2-13b",
+    "Sao10K/Medusa-13b",
+    "OptimalScale/robin-65b-v2-delta",
+    "minlik/chinese-alpaca-33b-merged",
+    "OpenAssistant/llama2-13b-megacode2-oasst",
+    "TheBloke/OpenAssistant-SFT-7-Llama-30B-HF",
+    "Undi95/UndiMix-v1-13b",
+    "ehartford/Samantha-1.11-13b",
+    "beaugogh/Llama2-13b-sharegpt4",
+    "Aeala/GPT4-x-AlpacaDente2-30b",
+    "luffycodes/nash-vicuna-13b-v1dot5-ep2-w-rag-w-simple",
+    "WizardLM/WizardLM-13B-V1.1",
+    "uukuguy/speechless-orca-platypus-coig-lite-2k-0.6e-13b",
+    "huggyllama/llama-30b",
+    "Undi95/ReMM-L2-13B-PIPPA",
+    "Undi95/ReMM-L2-13B",
+    "gaodrew/gaodrew-gorgonzola-13b",
+    "lmsys/vicuna-13b-v1.5",
+    "yeontaek/Platypus2xOpenOrca-13B-LoRa",
+    "Yhyu13/llama-30B-hf-openassitant",
+    "huggingface/llama-30b",
+    "lmsys/vicuna-13b-v1.5",
+    "TFLai/Athena-Platypus2-13B-QLora-0.80-epoch",
+    "TheBloke/dromedary-65b-lora-HF",
+    "yeontaek/llama-2-13b-Beluga-QLoRA",
+    "The-Face-Of-Goonery/Huginn-13b-V4",
+    "The-Face-Of-Goonery/Huginn-13b-v4.5",
+    "The-Face-Of-Goonery/Huginn-v3-13b",
+    "tiiuae/falcon-40b",
+    "WhoTookMyAmogusNickname/NewHope_HF_not_official",
+    "gaodrew/OpenOrca-Platypus2-13B-thera-1250",
+    "SLAM-group/NewHope",
+    "garage-bAInd/Platypus2-13B",
+    "migtissera/Synthia-13B",
+    "elinas/chronos-13b-v2",
+    "mosaicml/mpt-30b-chat",
+    "CHIH-HUNG/llama-2-13b-OpenOrca_5w",
+    "uukuguy/speechless-hermes-coig-lite-13b",
+    "TheBloke/tulu-30B-fp16",
+    "uukuguy/speechless-hermes-coig-lite-13b",
+    "xDAN-AI/xDAN_13b_l2_lora",
+    "lmsys/vicuna-13b-v1.5-16k",
+    "openchat/openchat_v3.1",
+    "CHIH-HUNG/llama-2-13b-dolphin_5w",
+    "Aspik101/vicuna-13b-v1.5-PL-lora_unload",
+    "Undi95/MLewd-L2-13B",
+    "ehartford/minotaur-llama2-13b-qlora",
+    "kajdun/iubaris-13b-v3",
+    "TFLai/Limarp-Platypus2-13B-QLoRA-0.80-epoch",
+    "openchat/openchat_v3.1",
+    "uukuguy/speechless-orca-platypus-coig-lite-4k-0.6e-13b",
+    "ziqingyang/chinese-alpaca-2-13b",
+    "TFLai/Airboros2.1-Platypus2-13B-QLora-0.80-epoch",
+    "yeontaek/llama-2-13b-Guanaco-QLoRA",
+    "lmsys/vicuna-13b-v1.5-16k",
+    "ehartford/based-30b",
+    "kingbri/airolima-chronos-grad-l2-13B",
+    "openchat/openchat_v3.2",
+    "uukuguy/speechless-orca-platypus-coig-lite-4k-0.5e-13b",
+    "yeontaek/Platypus2-13B-LoRa",
+    "kingbri/chronolima-airo-grad-l2-13B",
+    "openchat/openchat_v3.2",
+    "TFLai/PuddleJumper-Platypus2-13B-QLoRA-0.80-epoch",
+    "shareAI/llama2-13b-Chinese-chat",
+    "ehartford/WizardLM-1.0-Uncensored-Llama2-13b",
+    "Aspik101/Redmond-Puffin-13B-instruct-PL-lora_unload",
+    "yeontaek/llama-2-13B-ensemble-v6",
+    "WizardLM/WizardLM-13B-V1.2",
+    "TheBloke/WizardLM-13B-V1.1-GPTQ",
+    "bhenrym14/airophin-13b-pntk-16k-fp16",
+    "ehartford/WizardLM-1.0-Uncensored-Llama2-13b",
+    "Mikael110/llama-2-13b-guanaco-fp16",
+    "yeontaek/airoboros-2.1-llama-2-13B-QLoRa",
+    "CalderaAI/13B-Legerdemain-L2",
+    "grimpep/llama2-22b-wizard_vicuna",
+    "grimpep/llama2-22B-GPLATTY",
+    "bhenrym14/airophin-13b-pntk-16k-fp16",
+    "yeontaek/llama-2-13b-QLoRA",
+    "OpenAssistant/llama2-13b-orca-8k-3319",
+    "TheBloke/WizardLM-13B-V1-1-SuperHOT-8K-fp16",
+    "duliadotio/dulia-13b-8k-alpha",
+    "Undi95/LewdEngine",
+    "OpenBuddy/openbuddy-llama2-13b-v8.1-fp16",
+    "CHIH-HUNG/llama-2-13b-open_orca_20w",
+    "bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16",
+    "FlagAlpha/Llama2-Chinese-13b-Chat",
+    "LLMs/WizardLM-13B-V1.0",
+    "chansung/gpt4-alpaca-lora-13b-decapoda-1024",
+    "TheBloke/wizardLM-13B-1.0-fp16",
+    "digitous/13B-Chimera",
+    "yeontaek/Platypus2xOpenOrcaxGuanaco-13B-LoRa",
+    "jondurbin/airoboros-l2-13b-2.1",
+    "Monero/WizardLM-30B-Uncensored-Guanaco-SuperCOT-30b",
+    "TheBloke/UltraLM-13B-fp16",
+    "openaccess-ai-collective/minotaur-13b-fixed",
+    "NousResearch/Redmond-Puffin-13B",
+    "KoboldAI/LLaMA2-13B-Holomax",
+    "Lajonbot/WizardLM-13B-V1.2-PL-lora_unload",
+    "yeontaek/Platypus2-13B-LoRa-v2",
+    "TheBloke/airoboros-13B-HF",
+    "jondurbin/airoboros-13b",
+    "jjaaaww/posi_13b",
+    "CoolWP/llama-2-13b-guanaco-fp16",
+    "yeontaek/Platypus2-13B-QLoRa",
+    "h2oai/h2ogpt-research-oig-oasst1-512-30b",
+    "dfurman/llama-2-13b-guanaco-peft",
+    "NousResearch/Redmond-Puffin-13B",
+    "pe-nlp/llama-2-13b-platypus-vicuna-wizard",
+    "CHIH-HUNG/llama-2-13b-dolphin_20w",
+    "NousResearch/Nous-Hermes-13b",
+    "NobodyExistsOnTheInternet/GiftedConvo13bLoraNoEconsE4",
+    "ehartford/Wizard-Vicuna-13B-Uncensored",
+    "TheBloke/Wizard-Vicuna-13B-Uncensored-HF",
+    "openchat/openchat_v3.2_super",
+    "bhenrym14/airophin-v2-13b-PI-8k-fp16",
+    "openaccess-ai-collective/manticore-13b",
+    "The-Face-Of-Goonery/Huginn-22b-Prototype",
+    "jphme/Llama-2-13b-chat-german",
+    "grimpep/llama2-28B-Airo03",
+    "TheBloke/Kimiko-v2-13B-fp16",
+    "FPHam/Free_Sydney_13b_HF",
+    "lmsys/vicuna-13b-v1.3",
+    "FelixChao/llama2-13b-math1.1",
+    "CalderaAI/13B-BlueMethod",
+    "meta-llama/Llama-2-13b-chat-hf",
+    "deepse/CodeUp-Llama-2-13b-chat-hf",
+    "WizardLM/WizardMath-13B-V1.0",
+    "WizardLM/WizardMath-13B-V1.0",
+    "HyperbeeAI/Tulpar-7b-v0",
+    "xxyyy123/test_qkvo_adptor",
+    "xxyyy123/mc_data_30k_from_platpus_orca_7b_10k_v1_lora_qkvo_rank14_v2",
+    "openchat/openchat_v2_w",
+    "FelixChao/llama2-13b-math1.1",
+    "psmathur/orca_mini_v3_7b",
+    "TehVenom/Metharme-13b-Merged",
+    "xxyyy123/10k_v1_lora_qkvo_rank14_v3",
+    "OpenAssistant/llama2-13b-orca-v2-8k-3166",
+    "openaccess-ai-collective/wizard-mega-13b",
+    "jondurbin/airoboros-13b-gpt4-1.4",
+    "jondurbin/airoboros-13b-gpt4-1.4-fp16",
+    "Monero/Manticore-13b-Chat-Pyg-Guanaco",
+    "FelixChao/llama2-13b-math1.2",
+    "chargoddard/platypus-2-22b-relora",
+    "FelixChao/llama2-13b-math1.2",
+    "Gryphe/MythoBoros-13b",
+    "CalderaAI/13B-Ouroboros",
+    "OpenAssistant/llama2-13b-orca-v2-8k-3166",
+    "heegyu/LIMA2-13b-hf",
+    "digitous/13B-HyperMantis",
+    "Gryphe/MythoLogic-13b",
+    "TheBloke/Airoboros-L2-13B-2.1-GPTQ",
+    "chargoddard/platypus2-22b-relora",
+    "openchat/openchat_v2",
+    "yeontaek/Platypus2-13B-IA3",
+    "stabilityai/StableBeluga-7B",
+    "circulus/Llama-2-7b-orca-v1",
+    "budecosystem/genz-13b-v2",
+    "TheBloke/gpt4-x-vicuna-13B-HF",
+    "NobodyExistsOnTheInternet/GiftedConvo13bLoraNoEcons",
+    "zarakiquemparte/zarafusionex-1.1-l2-7b",
+    "Lajonbot/tableBeluga-7B-instruct-pl-lora_unload",
+    "jondurbin/airoboros-13b-gpt4",
+    "gaodrew/gaodrew-gorgonzola-13b",
+    "jondurbin/airoboros-13b-gpt4-1.1",
+    "TheBloke/gpt4-alpaca-lora-13B-HF",
+    "zarakiquemparte/zarablendex-vq-l2-7b",
+    "openaccess-ai-collective/manticore-13b-chat-pyg",
+    "Lajonbot/Llama-2-13b-hf-instruct-pl-lora_unload",
+    "NobodyExistsOnTheInternet/PuffedLIMA13bQLORA",
+    "xxyyy123/10k_v1_lora_qkvo_rank28_v2",
+    "jondurbin/airoboros-l2-13b-gpt4-1.4.1",
+    "dhmeltzer/Llama-2-13b-hf-eli5-wiki-1024_r_64_alpha_16",
+    "NobodyExistsOnTheInternet/PuffedConvo13bLoraE4",
+    "yihan6324/llama2-7b-instructmining-40k-sharegpt",
+    "CHIH-HUNG/llama-2-13b-Open_Platypus_and_ccp_2.6w",
+    "Aeala/GPT4-x-Alpasta-13b",
+    "psmathur/orca_mini_v2_13b",
+    "YeungNLP/firefly-llama-13b",
+    "psmathur/orca_mini_v2_13b",
+    "zarakiquemparte/zarafusionix-l2-7b",
+    "yihan6324/llama2-7b-instructmining-60k-sharegpt",
+    "yihan6324/llama-2-7b-instructmining-60k-sharegpt",
+    "layoric/llama-2-13b-code-alpaca",
+    "bofenghuang/vigogne-13b-instruct",
+    "Lajonbot/vicuna-13b-v1.3-PL-lora_unload",
+    "lvkaokao/llama2-7b-hf-chat-lora-v3",
+    "ehartford/dolphin-llama-13b",
+    "YeungNLP/firefly-llama-13b-v1.2",
+    "TheBloke/Kimiko-13B-fp16",
+    "kevinpro/Vicuna-13B-CoT",
+    "eachadea/vicuna-13b-1.1",
+    "pillowtalks-ai/delta13b",
+    "TheBloke/vicuna-13B-1.1-HF",
+    "TheBloke/Vicuna-13B-CoT-fp16",
+    "lmsys/vicuna-13b-delta-v1.1",
+    "lmsys/vicuna-13b-v1.1",
+    "xxyyy123/20k_v1_lora_qkvo_rank14_v2",
+    "TheBloke/guanaco-13B-HF",
+    "TheBloke/vicuna-13b-v1.3.0-GPTQ",
+    "edor/Stable-Platypus2-mini-7B",
+    "totally-not-an-llm/EverythingLM-13b-V2-16k",
+    "zarakiquemparte/zaraxe-l2-7b",
+    "beaugogh/Llama2-7b-openorca-mc-v2",
+    "TheBloke/Nous-Hermes-13B-SuperHOT-8K-fp16",
+    "quantumaikr/QuantumLM",
+    "jondurbin/airoboros-13b-gpt4-1.2",
+    "TheBloke/robin-13B-v2-fp16",
+    "TFLai/llama-2-13b-4bit-alpaca-gpt4",
+    "yihan6324/llama2-7b-instructmining-orca-40k",
+    "dvruette/oasst-llama-13b-2-epochs",
+    "Open-Orca/LlongOrca-7B-16k",
+    "Aspik101/Nous-Hermes-13b-pl-lora_unload",
+    "ehartford/Samantha-1.11-CodeLlama-34b",
+    "nkpz/llama2-22b-chat-wizard-uncensored",
+    "bofenghuang/vigogne-13b-chat",
+    "beaugogh/Llama2-7b-openorca-mc-v1",
+    "OptimalScale/robin-13b-v2-delta",
+    "pe-nlp/llama-2-13b-vicuna-wizard",
+    "chargoddard/llama2-22b",
+    "gywy/llama2-13b-chinese-v1",
+    "frank098/Wizard-Vicuna-13B-juniper",
+    "IGeniusDev/llama13B-quant8-testv1-openorca-customdataset",
+    "CHIH-HUNG/llama-2-13b-huangyt_Fintune_1_17w-gate_up_down_proj",
+    "eachadea/vicuna-13b",
+    "yihan6324/llama2-7b-instructmining-orca-90k",
+    "chargoddard/llama2-22b-blocktriangular",
+    "luffycodes/mcq-vicuna-13b-v1.5",
+    "Yhyu13/chimera-inst-chat-13b-hf",
+    "luffycodes/mcq-vicuna-13b-v1.5",
+    "chargoddard/ypotryll-22b-epoch2-qlora",
+    "totally-not-an-llm/EverythingLM-13b-16k",
+    "luffycodes/mcq-hal-vicuna-13b-v1.5",
+    "openaccess-ai-collective/minotaur-13b",
+    "IGeniusDev/llama13B-quant8-testv1-openorca-customdataset",
+    "chargoddard/llama2-22b-blocktriangular",
+    "TFLai/Platypus2-13B-QLoRA-0.80-epoch",
+    "meta-llama/Llama-2-13b-hf",
+    "CHIH-HUNG/llama-2-13b-huangyt_FINETUNE2_3w-gate_up_down_proj",
+    "luffycodes/mcq-hal-vicuna-13b-v1.5",
+    "TheBloke/Llama-2-13B-fp16",
+    "TaylorAI/Flash-Llama-13B",
+    "shareAI/bimoGPT-llama2-13b",
+    "wahaha1987/llama_13b_sharegpt94k_fastchat",
+    "openchat/openchat_8192",
+    "CHIH-HUNG/llama-2-13b-huangyt_Fintune_1_17w-q_k_v_o_proj",
+    "dvruette/llama-13b-pretrained-sft-do2",
+    "CHIH-HUNG/llama-2-13b-alpaca-test",
+    "OpenBuddy/openbuddy-llama2-13b-v11.1-bf16",
+    "CHIH-HUNG/llama-2-13b-FINETUNE2_TEST_2.2w",
+    "project-baize/baize-v2-13b",
+    "jondurbin/airoboros-l2-13b-gpt4-m2.0",
+    "yeontaek/Platypus2xOpenOrca-13B-LoRa-v2",
+    "CHIH-HUNG/llama-2-13b-huangyt_FINETUNE2_3w",
+    "xzuyn/Alpacino-SuperCOT-13B",
+    "jondurbin/airoboros-l2-13b-gpt4-2.0",
+    "aiplanet/effi-13b",
+    "clibrain/Llama-2-13b-ft-instruct-es",
+    "CHIH-HUNG/llama-2-13b-huangyt_Fintune_1_17w",
+    "bofenghuang/vigogne-2-7b-instruct",
+    "CHIH-HUNG/llama-2-13b-huangyt_FINETUNE2_3w-q_k_v_o_proj",
+    "bofenghuang/vigogne-2-7b-chat",
+    "aiplanet/effi-13b",
+    "haonan-li/bactrian-x-llama-13b-merged",
+    "beaugogh/Llama2-7b-sharegpt4",
+    "HWERI/Llama2-7b-sharegpt4",
+    "jondurbin/airoboros-13b-gpt4-1.3",
+    "jondurbin/airoboros-c34b-2.1",
+    "junelee/wizard-vicuna-13b",
+    "TheBloke/wizard-vicuna-13B-HF",
+    "Open-Orca/OpenOrca-Preview1-13B",
+    "TheBloke/h2ogpt-oasst1-512-30B-HF",
+    "TheBloke/Llama-2-13B-GPTQ",
+    "camel-ai/CAMEL-13B-Combined-Data",
+    "lmsys/vicuna-7b-v1.5",
+    "lmsys/vicuna-7b-v1.5-16k",
+    "lmsys/vicuna-7b-v1.5",
+    "ausboss/llama-13b-supercot",
+    "TheBloke/tulu-13B-fp16",
+    "NousResearch/Nous-Hermes-llama-2-7b",
+    "jlevin/guanaco-13b-llama-2",
+    "lmsys/vicuna-7b-v1.5-16k",
+    "dvruette/llama-13b-pretrained",
+    "nkpz/llama2-22b-daydreamer-v3",
+    "dvruette/llama-13b-pretrained-dropout",
+    "jondurbin/airoboros-l2-13b-2.1",
+    "LLMs/Stable-Vicuna-13B",
+    "64bits/LexPodLM-13B",
+    "lizhuang144/llama_mirror_13b_v1.0",
+    "TheBloke/stable-vicuna-13B-HF",
+    "zarakiquemparte/zaraxls-l2-7b",
+    "TheBloke/Llama-2-13B-GPTQ",
+    "Kiddyz/testlm-3",
+    "migtissera/Synthia-7B",
+    "zarakiquemparte/zarablend-l2-7b",
+    "mosaicml/mpt-30b-instruct",
+    "PocketDoc/Dans-PileOfSets-Mk1-llama-13b-merged",
+    "vonjack/Qwen-LLaMAfied-HFTok-7B-Chat",
+    "l3utterfly/llama2-7b-layla",
+    "Lajonbot/vicuna-7b-v1.5-PL-lora_unload",
+    "heegyu/LIMA-13b-hf",
+    "frank098/WizardLM_13B_juniper",
+    "ashercn97/manatee-7b",
+    "chavinlo/gpt4-x-alpaca",
+    "PocketDoc/Dans-PersonalityEngine-13b",
+    "ehartford/WizardLM-1.0-Uncensored-CodeLlama-34b",
+    "digitous/Alpacino13b",
+    "edor/Hermes-Platypus2-mini-7B",
+    "lvkaokao/llama2-7b-hf-chat-lora-v2",
+    "Kiddyz/testlm-1-1",
+    "Kiddyz/testlm",
+    "Kiddyz/testlm-1",
+    "Kiddyz/testlm2",
+    "radm/Philosophy-Platypus2-13b",
+    "aiplanet/effi-13b",
+    "Harshvir/Llama-2-7B-physics",
+    "YeungNLP/firefly-ziya-13b",
+    "LinkSoul/Chinese-Llama-2-7b",
+    "PeanutJar/LLaMa-2-PeanutButter_v10-7B",
+    "OpenBuddy/openbuddy-llama2-13b-v11-bf16",
+    "StudentLLM/Alpagasus-2-13B-QLoRA-pipeline",
+    "meta-llama/Llama-2-13b-hf",
+    "WizardLM/WizardCoder-Python-34B-V1.0",
+    "dvruette/llama-13b-pretrained-sft-epoch-1",
+    "camel-ai/CAMEL-13B-Role-Playing-Data",
+    "ziqingyang/chinese-llama-2-13b",
+    "rombodawg/LosslessMegaCoder-llama2-7b-mini",
+    "TheBloke/koala-13B-HF",
+    "lmsys/vicuna-7b-delta-v1.1",
+    "eachadea/vicuna-7b-1.1",
+    "Ejafa/vicuna_7B_vanilla_1.1",
+    "lvkaokao/llama2-7b-hf-chat-lora",
+    "OpenBuddy/openbuddy-atom-13b-v9-bf16",
+    "Norquinal/llama-2-7b-claude-chat-rp",
+    "Danielbrdz/Barcenas-7b",
+    "heegyu/WizardVicuna2-13b-hf",
+    "meta-llama/Llama-2-7b-chat-hf",
+    "PeanutJar/LLaMa-2-PeanutButter_v14-7B",
+    "PeanutJar/LLaMa-2-PeanutButter_v4-7B",
+    "davzoku/cria-llama2-7b-v1.3",
+    "OpenBuddy/openbuddy-atom-13b-v9-bf16",
+    "lvkaokao/llama2-7b-hf-instruction-lora",
+    "Tap-M/Luna-AI-Llama2-Uncensored",
+    "ehartford/Samantha-1.11-7b",
+    "WizardLM/WizardCoder-Python-34B-V1.0",
+    "TheBloke/Manticore-13B-Chat-Pyg-Guanaco-SuperHOT-8K-GPTQ",
+    "Mikael110/llama-2-7b-guanaco-fp16",
+    "garage-bAInd/Platypus2-7B",
+    "PeanutJar/LLaMa-2-PeanutButter_v18_B-7B",
+    "mosaicml/mpt-30b",
+    "garage-bAInd/Platypus2-7B",
+    "huggingface/llama-13b",
+    "dvruette/oasst-llama-13b-1000-steps",
+    "jordiclive/gpt4all-alpaca-oa-codealpaca-lora-13b",
+    "huggyllama/llama-13b",
+    "Voicelab/trurl-2-7b",
+    "TFLai/llama-13b-4bit-alpaca",
+    "gywy/llama2-13b-chinese-v2",
+    "lmsys/longchat-13b-16k",
+    "Aspik101/trurl-2-7b-pl-instruct_unload",
+    "WizardLM/WizardMath-7B-V1.0",
+    "Norquinal/llama-2-7b-claude-chat",
+    "TheTravellingEngineer/llama2-7b-chat-hf-dpo",
+    "HuggingFaceH4/starchat-beta",
+    "joehuangx/spatial-vicuna-7b-v1.5-LoRA",
+    "conceptofmind/LLongMA-2-13b-16k",
+    "tianyil1/denas-llama2",
+    "lmsys/vicuna-7b-v1.3",
+    "conceptofmind/LLongMA-2-13b-16k",
+    "openchat/opencoderplus",
+    "ajibawa-2023/scarlett-7b",
+    "dhmeltzer/llama-7b-SFT_eli5_wiki65k_1024_r_64_alpha_16_merged",
+    "psyche/kollama2-7b-v2",
+    "heegyu/LIMA2-7b-hf",
+    "dhmeltzer/llama-7b-SFT-qlora-eli5-wiki_DPO_ds_RM_top_2_1024_r_64_alpha_16",
+    "abhishek/llama2guanacotest",
+    "jondurbin/airoboros-l2-7b-2.1",
+    "llama-anon/instruct-13b",
+    "FelixChao/vicuna-7B-physics",
+    "Aspik101/Llama-2-7b-hf-instruct-pl-lora_unload",
+    "shibing624/chinese-alpaca-plus-13b-hf",
+    "davzoku/cria-llama2-7b-v1.3_peft",
+    "quantumaikr/llama-2-7b-hf-guanaco-1k",
+    "togethercomputer/Llama-2-7B-32K-Instruct",
+    "sia-ai/llama-2-7b-1-percent-open-orca-1000-steps-v0",
+    "TheTravellingEngineer/llama2-7b-hf-guanaco",
+    "Lajonbot/Llama-2-7b-chat-hf-instruct-pl-lora_unload",
+    "jondurbin/airoboros-l2-7b-gpt4-1.4.1",
+    "wahaha1987/llama_7b_sharegpt94k_fastchat",
+    "FelixChao/vicuna-7B-chemical",
+    "TinyPixel/llama2-7b-oa",
+    "chaoyi-wu/MedLLaMA_13B",
+    "edor/Platypus2-mini-7B",
+    "RoversX/llama-2-7b-hf-small-shards-Samantha-V1-SFT",
+    "venkycs/llama-v2-7b-32kC-Security",
+    "psyche/kollama2-7b",
+    "Fredithefish/Guanaco-7B-Uncensored",
+    "TheTravellingEngineer/llama2-7b-chat-hf-guanaco",
+    "ehartford/WizardLM-13B-Uncensored",
+    "PocketDoc/Dans-CreepingSenseOfDoom",
+    "wenge-research/yayi-7b-llama2",
+    "georgesung/llama2_7b_chat_uncensored",
+    "TinyPixel/llama2-7b-instruct",
+    "quantumaikr/QuantumLM-7B",
+    "xzuyn/MedicWizard-7B",
+    "wenge-research/yayi-7b-llama2",
+    "TinyPixel/lima-test",
+    "elyza/ELYZA-japanese-Llama-2-7b-instruct",
+    "lgaalves/llama-2-7b-hf_open-platypus",
+    "ziqingyang/chinese-alpaca-2-7b",
+    "TehVenom/Pygmalion-Vicuna-1.1-7b",
+    "meta-llama/Llama-2-7b-hf",
+    "bongchoi/test-llama2-7b",
+    "TaylorAI/Flash-Llama-7B",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v2",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v4",
+    "kashif/stack-llama-2",
+    "PeanutJar/LLaMa-2-PeanutButter_v18_A-7B",
+    "ToolBench/ToolLLaMA-7b-LoRA",
+    "Monero/WizardLM-13b-OpenAssistant-Uncensored",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v2",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v4",
+    "mrm8488/llama-2-coder-7b",
+    "elyza/ELYZA-japanese-Llama-2-7b-fast-instruct",
+    "clibrain/Llama-2-7b-ft-instruct-es",
+    "medalpaca/medalpaca-7b",
+    "TheBloke/tulu-7B-fp16",
+    "OpenBuddy/openbuddy-openllama-13b-v7-fp16",
+    "TaylorAI/FLAN-Llama-7B-2_Llama2-7B-Flash_868_full_model",
+    "Aspik101/vicuna-7b-v1.3-instruct-pl-lora_unload",
+    "jondurbin/airoboros-l2-7b-gpt4-2.0",
+    "dhmeltzer/llama-7b-SFT_ds_eli5_1024_r_64_alpha_16_merged",
+    "GOAT-AI/GOAT-7B-Community",
+    "AtomEchoAI/AtomGPT_56k",
+    "julianweng/Llama-2-7b-chat-orcah",
+    "TehVenom/Pygmalion-13b-Merged",
+    "jondurbin/airoboros-7b-gpt4-1.1",
+    "dhmeltzer/llama-7b-SFT_ds_wiki65k_1024_r_64_alpha_16_merged",
+    "bofenghuang/vigogne-7b-chat",
+    "lmsys/longchat-7b-v1.5-32k",
+    "jondurbin/airoboros-l2-7b-gpt4-m2.0",
+    "synapsoft/Llama-2-7b-chat-hf-flan2022-1.2M",
+    "jondurbin/airoboros-7b-gpt4-1.4",
+    "Charlie911/vicuna-7b-v1.5-lora-mctaco",
+    "yihan6324/instructmining-platypus-15k",
+    "meta-llama/Llama-2-7b-hf",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v3",
+    "quantumaikr/KoreanLM-hf",
+    "openthaigpt/openthaigpt-1.0.0-alpha-7b-chat-ckpt-hf",
+    "TheBloke/Llama-2-7B-GPTQ",
+    "TheBloke/Llama-2-7B-GPTQ",
+    "LLMs/AlpacaGPT4-7B-elina",
+    "ehartford/Wizard-Vicuna-7B-Uncensored",
+    "TheBloke/Wizard-Vicuna-7B-Uncensored-HF",
+    "TheTravellingEngineer/llama2-7b-chat-hf-v3",
+    "golaxy/gowizardlm",
+    "ehartford/dolphin-llama2-7b",
+    "CHIH-HUNG/llama-2-7b-dolphin_10w-test",
+    "mncai/chatdoctor",
+    "psyche/kollama2-7b-v3",
+    "jondurbin/airoboros-7b-gpt4",
+    "jondurbin/airoboros-7b",
+    "TheBloke/airoboros-7b-gpt4-fp16",
+    "mosaicml/mpt-7b-8k-chat",
+    "elyza/ELYZA-japanese-Llama-2-7b",
+    "bofenghuang/vigogne-7b-instruct",
+    "jxhong/CAlign-alpaca-7b",
+    "golaxy/goims",
+    "jondurbin/airoboros-7b-gpt4-1.2",
+    "jphme/orca_mini_v2_ger_7b",
+    "psmathur/orca_mini_v2_7b",
+    "notstoic/PygmalionCoT-7b",
+    "golaxy/gogpt2-13b",
+    "golaxy/gogpt2-13b-chat",
+    "togethercomputer/LLaMA-2-7B-32K",
+    "TheBloke/wizardLM-7B-HF",
+    "keyfan/vicuna-chinese-replication-v1.1",
+    "golaxy/gogpt2-7b",
+    "aiplanet/effi-7b",
+    "arver/llama7b-qlora",
+    "titan087/OpenLlama13B-Guanaco",
+    "chavinlo/alpaca-native",
+    "project-baize/baize-healthcare-lora-7B",
+    "AlpinDale/pygmalion-instruct",
+    "openlm-research/open_llama_13b",
+    "jondurbin/airoboros-7b-gpt4-1.3",
+    "elyza/ELYZA-japanese-Llama-2-7b-fast",
+    "jondurbin/airoboros-gpt-3.5-turbo-100k-7b",
+    "uukuguy/speechless-codellama-orca-13b",
+    "bigcode/starcoderplus",
+    "TheBloke/guanaco-7B-HF",
+    "Neko-Institute-of-Science/metharme-7b",
+    "TigerResearch/tigerbot-7b-base",
+    "golaxy/gogpt-7b",
+    "togethercomputer/LLaMA-2-7B-32K",
+    "yhyhy3/open_llama_7b_v2_med_instruct",
+    "ajibawa-2023/carl-7b",
+    "stabilityai/stablelm-base-alpha-7b-v2",
+    "conceptofmind/LLongMA-2-7b-16k",
+    "TehVenom/Pygmalion_AlpacaLora-7b",
+    "jondurbin/airoboros-7b-gpt4-1.4.1-qlora",
+    "wannaphong/openthaigpt-0.1.0-beta-full-model_for_open_llm_leaderboard",
+    "ausboss/llama7b-wizardlm-unfiltered",
+    "project-baize/baize-v2-7b",
+    "LMFlow/Robin-v2",
+    "HanningZhang/Robin-v2",
+    "LMFlow/Robin-7b-v2",
+    "OptimalScale/robin-7b-v2-delta",
+    "uukuguy/speechless-codellama-platypus-13b",
+    "jerryjalapeno/nart-100k-7b",
+    "wenge-research/yayi-13b-llama2",
+    "fireballoon/baichuan-vicuna-chinese-7b",
+    "jlevin/guanaco-unchained-llama-2-7b",
+    "csitfun/llama-7b-logicot",
+    "DevaMalla/llama7b_alpaca_1gpu_bf16",
+    "WeOpenML/PandaLM-Alpaca-7B-v1",
+    "illuin/test-custom-llama",
+    "yeontaek/WizardCoder-Python-13B-LoRa",
+    "ashercn97/giraffe-7b",
+    "mosaicml/mpt-7b-chat",
+    "abhishek/autotrain-llama-alpaca-peft-52508123785",
+    "Neko-Institute-of-Science/pygmalion-7b",
+    "TFLai/llama-7b-4bit-alpaca",
+    "huggingface/llama-7b",
+    "TheBloke/Planner-7B-fp16",
+    "shibing624/chinese-llama-plus-13b-hf",
+    "AGI-inc/lora_moe_7b_baseline",
+    "DevaMalla/llama-base-7b",
+    "AGI-inc/lora_moe_7b",
+    "togethercomputer/GPT-JT-6B-v0",
+    "ehartford/WizardLM-7B-Uncensored",
+    "shibing624/chinese-alpaca-plus-7b-hf",
+    "beomi/llama-2-ko-7b",
+    "mosaicml/mpt-7b-8k-instruct",
+    "Enno-Ai/ennodata-7b",
+    "mosaicml/mpt-7b-instruct",
+    "facebook/opt-iml-max-30b",
+    "WeOpenML/Alpaca-7B-v1",
+    "TheBloke/Project-Baize-v2-7B-GPTQ",
+    "codellama/CodeLlama-13b-Instruct-hf",
+    "TheBloke/CodeLlama-13B-Instruct-fp16",
+    "facebook/galactica-30b",
+    "FreedomIntelligence/phoenix-inst-chat-7b",
+    "openlm-research/open_llama_7b_v2",
+    "GeorgiaTechResearchInstitute/galpaca-30b",
+    "THUDM/chatglm2-6b",
+    "togethercomputer/GPT-JT-6B-v1",
+    "TheBloke/koala-7B-HF",
+    "nathan0/mpt_delta_tuned_model_v3",
+    "nathan0/mpt_delta_tuned_model_v2",
+    "GeorgiaTechResearchInstitute/galpaca-30b",
+    "JosephusCheung/Guanaco",
+    "shareAI/CodeLLaMA-chat-13b-Chinese",
+    "TigerResearch/tigerbot-7b-sft",
+    "Writer/InstructPalmyra-20b",
+    "OpenAssistant/codellama-13b-oasst-sft-v10",
+    "bigscience/bloomz-7b1-mt",
+    "nathan0/mpt_delta_tuned_model_v3",
+    "VMware/open-llama-7b-open-instruct",
+    "baichuan-inc/Baichuan-7B",
+    "anas-awadalla/mpt-7b",
+    "mosaicml/mpt-7b",
+    "bigscience/bloomz-7b1",
+    "ziqingyang/chinese-llama-2-7b",
+    "OpenAssistant/codellama-13b-oasst-sft-v10",
+    "wenge-research/yayi-7b",
+    "tiiuae/falcon-7b",
+    "togethercomputer/RedPajama-INCITE-Instruct-7B-v0.1",
+    "togethercomputer/RedPajama-INCITE-7B-Instruct",
+    "TheBloke/landmark-attention-llama7b-fp16",
+    "togethercomputer/GPT-JT-Moderation-6B",
+    "h2oai/h2ogpt-gm-oasst1-en-1024-20b",
+    "dvruette/gpt-neox-20b-full-precision",
+    "TehVenom/Moderator-Chan_GPT-JT-6b",
+    "dvruette/oasst-gpt-neox-20b-1000-steps",
+    "AlekseyKorshuk/pygmalion-6b-vicuna-chatml",
+    "facebook/opt-66b",
+    "Salesforce/codegen-16B-nl",
+    "Vmware/open-llama-7b-v2-open-instruct",
+    "mosaicml/mpt-7b-storywriter",
+    "acrastt/Marx-3B-V2",
+    "openlm-research/open_llama_7b",
+    "Fredithefish/ReasonixPajama-3B-HF",
+    "togethercomputer/GPT-NeoXT-Chat-Base-20B",
+    "psmathur/orca_mini_13b",
+    "RWKV/rwkv-raven-14b",
+    "h2oai/h2ogpt-oasst1-512-20b",
+    "acrastt/Marx-3B",
+    "klosax/open_llama_13b_600bt_preview",
+    "synapsoft/Llama-2-7b-hf-flan2022-1.2M",
+    "OpenAssistant/oasst-sft-1-pythia-12b",
+    "golaxy/gogpt-7b-bloom",
+    "Writer/palmyra-large",
+    "psmathur/orca_mini_7b",
+    "dvruette/oasst-pythia-12b-6000-steps",
+    "NousResearch/CodeLlama-13b-hf",
+    "codellama/CodeLlama-13b-hf",
+    "h2oai/h2ogpt-gm-oasst1-multilang-1024-20b",
+    "VMware/open-llama-0.7T-7B-open-instruct-v1.1",
+    "dvruette/oasst-pythia-12b-flash-attn-5000-steps",
+    "dvruette/oasst-gpt-neox-20b-3000-steps",
+    "RobbeD/OpenLlama-Platypus-3B",
+    "facebook/opt-30b",
+    "acrastt/Puma-3B",
+    "OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
+    "dvruette/oasst-pythia-12b-pretrained-sft",
+    "digitous/GPT-R",
+    "acrastt/Griffin-3B",
+    "togethercomputer/RedPajama-INCITE-Base-7B-v0.1",
+    "togethercomputer/RedPajama-INCITE-7B-Base",
+    "CobraMamba/mamba-gpt-3b-v3",
+    "Danielbrdz/CodeBarcenas-7b",
+    "l3utterfly/open-llama-3b-v2-layla",
+    "CobraMamba/mamba-gpt-3b-v2",
+    "OpenAssistant/pythia-12b-sft-v8-7k-steps",
+    "KoboldAI/GPT-NeoX-20B-Erebus",
+    "RobbeD/Orca-Platypus-3B",
+    "h2oai/h2ogpt-gm-oasst1-en-1024-12b",
+    "OpenAssistant/pythia-12b-sft-v8-2.5k-steps",
+    "AlekseyKorshuk/chatml-pyg-v1",
+    "togethercomputer/RedPajama-INCITE-Chat-7B-v0.1",
+    "togethercomputer/RedPajama-INCITE-7B-Chat",
+    "digitous/Javelin-R",
+    "dvruette/oasst-pythia-12b-reference",
+    "EleutherAI/gpt-neox-20b",
+    "KoboldAI/fairseq-dense-13B",
+    "OpenAssistant/pythia-12b-sft-v8-rlhf-2k-steps",
+    "codellama/CodeLlama-7b-Instruct-hf",
+    "digitous/Javelin-GPTJ",
+    "KoboldAI/GPT-NeoX-20B-Skein",
+    "digitous/Javalion-R",
+    "h2oai/h2ogpt-oasst1-512-12b",
+    "acrastt/Bean-3B",
+    "KoboldAI/GPT-J-6B-Skein",
+    "nomic-ai/gpt4all-j",
+    "databricks/dolly-v2-12b",
+    "TehVenom/Dolly_Shygmalion-6b-Dev_V8P2",
+    "databricks/dolly-v2-7b",
+    "Aspik101/WizardVicuna-Uncensored-3B-instruct-PL-lora_unload",
+    "digitous/Adventien-GPTJ",
+    "openlm-research/open_llama_3b_v2",
+    "RWKV/rwkv-4-14b-pile",
+    "Lazycuber/Janemalion-6B",
+    "OpenAssistant/pythia-12b-pre-v8-12.5k-steps",
+    "digitous/Janin-R",
+    "kfkas/Llama-2-ko-7b-Chat",
+    "heegyu/WizardVicuna-Uncensored-3B-0719",
+    "h2oai/h2ogpt-gm-oasst1-en-1024-open-llama-7b-preview-400bt",
+    "TaylorAI/Flash-Llama-3B",
+    "kfkas/Llama-2-ko-7b-Chat",
+    "digitous/Skegma-GPTJ",
+    "digitous/Javalion-GPTJ",
+    "Pirr/pythia-13b-deduped-green_devil",
+    "TehVenom/PPO_Shygmalion-V8p4_Dev-6b",
+    "dvruette/oasst-pythia-6.9b-4000-steps",
+    "heegyu/WizardVicuna-3B-0719",
+    "psmathur/orca_mini_3b",
+    "OpenAssistant/galactica-6.7b-finetuned",
+    "frank098/orca_mini_3b_juniper",
+    "PygmalionAI/pygmalion-6b",
+    "TehVenom/PPO_Pygway-V8p4_Dev-6b",
+    "TFLai/gpt-neox-20b-4bit-alpaca",
+    "Corianas/gpt-j-6B-Dolly",
+    "TehVenom/Dolly_Shygmalion-6b",
+    "digitous/Janin-GPTJ",
+    "TehVenom/GPT-J-Pyg_PPO-6B-Dev-V8p4",
+    "EleutherAI/gpt-j-6b",
+    "KoboldAI/GPT-J-6B-Shinen",
+    "TehVenom/Dolly_Malion-6b",
+    "TehVenom/ChanMalion",
+    "Salesforce/codegen-6B-nl",
+    "Fredithefish/RedPajama-INCITE-Chat-3B-Instruction-Tuning-with-GPT-4",
+    "KoboldAI/GPT-J-6B-Janeway",
+    "togethercomputer/RedPajama-INCITE-Chat-3B-v1",
+    "togethercomputer/Pythia-Chat-Base-7B",
+    "heegyu/RedTulu-Uncensored-3B-0719",
+    "KoboldAI/PPO_Pygway-6b-Mix",
+    "KoboldAI/OPT-13B-Erebus",
+    "KoboldAI/fairseq-dense-6.7B",
+    "EleutherAI/pythia-12b-deduped",
+    "pszemraj/pythia-6.9b-HC3",
+    "Fredithefish/Guanaco-3B-Uncensored-v2",
+    "facebook/opt-13b",
+    "TehVenom/GPT-J-Pyg_PPO-6B",
+    "EleutherAI/pythia-6.9b-deduped",
+    "Devio/test-1400",
+    "Fredithefish/Guanaco-3B-Uncensored",
+    "codellama/CodeLlama-7b-hf",
+    "acrastt/RedPajama-INCITE-Chat-Instruct-3B-V1",
+    "Fredithefish/ScarletPajama-3B-HF",
+    "KoboldAI/OPT-13B-Nerybus-Mix",
+    "YeungNLP/firefly-bloom-7b1",
+    "DanielSc4/RedPajama-INCITE-Chat-3B-v1-RL-LoRA-8bit-test1",
+    "klosax/open_llama_7b_400bt_preview",
+    "KoboldAI/OPT-13B-Nerys-v2",
+    "TehVenom/PPO_Shygmalion-6b",
+    "amazon/LightGPT",
+    "KnutJaegersberg/black_goo_recipe_c",
+    "NousResearch/CodeLlama-7b-hf",
+    "togethercomputer/RedPajama-INCITE-Instruct-3B-v1",
+    "heegyu/WizardVicuna-open-llama-3b-v2",
+    "bigscience/bloom-7b1",
+    "Devio/test-22B",
+    "RWKV/rwkv-raven-7b",
+    "hakurei/instruct-12b",
+    "CobraMamba/mamba-gpt-3b",
+    "KnutJaegersberg/black_goo_recipe_a",
+    "acrastt/OmegLLaMA-3B",
+    "codellama/CodeLlama-7b-Instruct-hf",
+    "h2oai/h2ogpt-oig-oasst1-512-6_9b",
+    "KoboldAI/OPT-6.7B-Erebus",
+    "facebook/opt-6.7b",
+    "KnutJaegersberg/black_goo_recipe_d",
+    "KnutJaegersberg/LLongMA-3b-LIMA",
+    "KnutJaegersberg/black_goo_recipe_b",
+    "KoboldAI/OPT-6.7B-Nerybus-Mix",
+    "health360/Healix-3B",
+    "EleutherAI/pythia-12b",
+    "Fredithefish/RedPajama-INCITE-Chat-3B-ShareGPT-11K",
+    "GeorgiaTechResearchInstitute/galactica-6.7b-evol-instruct-70k",
+    "h2oai/h2ogpt-oig-oasst1-256-6_9b",
+    "ikala/bloom-zh-3b-chat",
+    "Taekyoon/llama2-ko-7b-test",
+    "anhnv125/pygmalion-6b-roleplay",
+    "TehVenom/DiffMerge_Pygmalion_Main-onto-V8P4",
+    "KoboldAI/OPT-6B-nerys-v2",
+    "Lazycuber/pyg-instruct-wizardlm",
+    "Devio/testC",
+    "KoboldAI/OPT-30B-Erebus",
+    "Fredithefish/CrimsonPajama",
+    "togethercomputer/RedPajama-INCITE-Base-3B-v1",
+    "bigscience/bloomz-3b",
+    "conceptofmind/Open-LLongMA-3b",
+    "RWKV/rwkv-4-7b-pile",
+    "openlm-research/open_llama_3b",
+    "ewof/koishi-instruct-3b",
+    "DanielSc4/RedPajama-INCITE-Chat-3B-v1-FT-LoRA-8bit-test1",
+    "cerebras/Cerebras-GPT-13B",
+    "EleutherAI/pythia-6.7b",
+    "aisquared/chopt-2_7b",
+    "Azure99/blossom-v1-3b",
+    "PSanni/Deer-3b",
+    "bertin-project/bertin-gpt-j-6B-alpaca",
+    "OpenBuddy/openbuddy-openllama-3b-v10-bf16",
+    "KoboldAI/fairseq-dense-2.7B",
+    "ehartford/CodeLlama-34b-Instruct-hf",
+    "codellama/CodeLlama-34b-Instruct-hf",
+    "TheBloke/CodeLlama-34B-Instruct-fp16",
+    "h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt-v2",
+    "openlm-research/open_llama_7b_700bt_preview",
+    "NbAiLab/nb-gpt-j-6B-alpaca",
+    "KoboldAI/OPT-2.7B-Erebus",
+    "Writer/camel-5b-hf",
+    "EleutherAI/pythia-2.7b",
+    "facebook/xglm-7.5B",
+    "EleutherAI/pythia-2.8b-deduped",
+    "klosax/open_llama_3b_350bt_preview",
+    "klosax/openllama-3b-350bt",
+    "KoboldAI/OPT-2.7B-Nerybus-Mix",
+    "KoboldAI/GPT-J-6B-Adventure",
+    "cerebras/Cerebras-GPT-6.7B",
+    "TFLai/pythia-2.8b-4bit-alpaca",
+    "facebook/opt-2.7b",
+    "KoboldAI/OPT-2.7B-Nerys-v2",
+    "bigscience/bloom-3b",
+    "Devio/test100",
+    "RWKV/rwkv-raven-3b",
+    "Azure99/blossom-v2-3b",
+    "codellama/CodeLlama-34b-Python-hf",
+    "bhenrym14/airoboros-33b-gpt4-1.4.1-PI-8192-fp16",
+    "EleutherAI/gpt-neo-2.7B",
+    "danielhanchen/open_llama_3b_600bt_preview",
+    "HuggingFaceH4/starchat-alpha",
+    "pythainlp/wangchanglm-7.5B-sft-en-sharded",
+    "beaugogh/pythia-1.4b-deduped-sharegpt",
+    "HWERI/pythia-1.4b-deduped-sharegpt",
+    "OpenAssistant/stablelm-7b-sft-v7-epoch-3",
+    "codellama/CodeLlama-7b-Python-hf",
+    "aisquared/chopt-1_3b",
+    "PygmalionAI/metharme-1.3b",
+    "Linly-AI/Chinese-LLaMA-2-13B-hf",
+    "chargoddard/llama-2-34b-uncode",
+    "RWKV/rwkv-4-3b-pile",
+    "pythainlp/wangchanglm-7.5B-sft-enth",
+    "MBZUAI/LaMini-GPT-1.5B",
+    "Writer/palmyra-base",
+    "KoboldAI/fairseq-dense-1.3B",
+    "EleutherAI/pythia-1.4b-deduped",
+    "MBZUAI/lamini-neo-1.3b",
+    "h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt",
+    "sartmis1/starcoder-finetune-openapi",
+    "MayaPH/opt-flan-iml-6.7b",
+    "facebook/xglm-4.5B",
+    "WizardLM/WizardCoder-15B-V1.0",
+    "facebook/opt-iml-max-1.3b",
+    "stabilityai/stablelm-tuned-alpha-7b",
+    "aisquared/dlite-v2-1_5b",
+    "stabilityai/stablelm-base-alpha-7b",
+    "sartmis1/starcoder-finetune-selfinstruct",
+    "lizhuang144/starcoder_mirror",
+    "bigcode/starcoder",
+    "TheBloke/CodeLlama-34B-Python-fp16",
+    "open-llm-leaderboard/bloomz-1b7-4bit-alpaca-auto-eval-adapter-applied",
+    "ehartford/CodeLlama-34b-Python-hf",
+    "codellama/CodeLlama-7b-Python-hf",
+    "GeorgiaTechResearchInstitute/starcoder-gpteacher-code-instruct",
+    "LoupGarou/WizardCoder-Guanaco-15B-V1.0",
+    "golaxy/gogpt-3b-bloom",
+    "EleutherAI/pythia-1.3b",
+    "codellama/CodeLlama-13b-Python-hf",
+    "hakurei/lotus-12B",
+    "NYTK/PULI-GPTrio",
+    "facebook/opt-1.3b",
+    "TheBloke/CodeLlama-13B-Python-fp16",
+    "codellama/CodeLlama-13b-Python-hf",
+    "RWKV/rwkv-raven-1b5",
+    "PygmalionAI/pygmalion-2.7b",
+    "bigscience/bloom-1b7",
+    "gpt2-xl",
+    "LoupGarou/WizardCoder-Guanaco-15B-V1.1",
+    "RWKV/rwkv-4-1b5-pile",
+    "codellama/CodeLlama-34b-hf",
+    "NousResearch/CodeLlama-34b-hf",
+    "rinna/bilingual-gpt-neox-4b-8k",
+    "lxe/Cerebras-GPT-2.7B-Alpaca-SP",
+    "cerebras/Cerebras-GPT-2.7B",
+    "jzjiao/opt-1.3b-rlhf",
+    "EleutherAI/gpt-neo-1.3B",
+    "aisquared/dlite-v1-1_5b",
+    "Corianas/Quokka_2.7b",
+    "MrNJK/gpt2-xl-sft",
+    "facebook/galactica-1.3b",
+    "aisquared/dlite-v2-774m",
+    "EleutherAI/pythia-1b-deduped",
+    "Kunhao/pile-7b-250b-tokens",
+    "w601sxs/b1ade-1b",
+    "rinna/bilingual-gpt-neox-4b",
+    "shaohang/SparseOPT-1.3B",
+    "shaohang/Sparse0.5_OPT-1.3",
+    "EleutherAI/polyglot-ko-12.8b",
+    "Salesforce/codegen-6B-multi",
+    "bigscience/bloom-1b1",
+    "TFLai/gpt-neo-1.3B-4bit-alpaca",
+    "FabbriSimo01/Bloom_1b_Quantized",
+    "MBZUAI/LaMini-GPT-774M",
+    "Locutusque/gpt2-large-conversational",
+    "Devio/test-3b",
+    "stabilityai/stablelm-tuned-alpha-3b",
+    "PygmalionAI/pygmalion-1.3b",
+    "KoboldAI/fairseq-dense-355M",
+    "Rachneet/gpt2-xl-alpaca",
+    "gpt2-large",
+    "Mikivis/gpt2-large-lora-sft",
+    "stabilityai/stablelm-base-alpha-3b",
+    "gpt2-medium",
+    "Kunhao/pile-7b",
+    "aisquared/dlite-v1-774m",
+    "aisquared/dlite-v2-355m",
+    "YeungNLP/firefly-bloom-2b6-v2",
+    "KnutJaegersberg/gpt-2-xl-EvolInstruct",
+    "KnutJaegersberg/galactica-orca-wizardlm-1.3b",
+    "cerebras/Cerebras-GPT-1.3B",
+    "FabbriSimo01/Cerebras_1.3b_Quantized",
+    "facebook/xglm-1.7B",
+    "EleutherAI/pythia-410m-deduped",
+    "TheBloke/GPlatty-30B-SuperHOT-8K-fp16",
+    "DataLinguistic/DataLinguistic-34B-V1.0",
+    "Corianas/Quokka_1.3b",
+    "TheTravellingEngineer/bloom-560m-RLHF-v2",
+    "Corianas/1.3b",
+    "RWKV/rwkv-4-430m-pile",
+    "porkorbeef/Llama-2-13b-sf",
+    "xhyi/PT_GPTNEO350_ATG",
+    "TheBloke/Wizard-Vicuna-13B-Uncensored-GPTQ",
+    "bigscience/bloomz-560m",
+    "TheBloke/medalpaca-13B-GPTQ-4bit",
+    "TheBloke/Vicuna-33B-1-3-SuperHOT-8K-fp16",
+    "aisquared/dlite-v1-355m",
+    "uukuguy/speechless-codellama-orca-airoboros-13b-0.10e",
+    "yhyhy3/med-orca-instruct-33b",
+    "TheBloke/Wizard-Vicuna-30B-Superhot-8K-fp16",
+    "TheTravellingEngineer/bloom-1b1-RLHF",
+    "MBZUAI/lamini-cerebras-1.3b",
+    "IDEA-CCNL/Ziya-LLaMA-13B-Pretrain-v1",
+    "TheBloke/WizardLM-7B-uncensored-GPTQ",
+    "TheBloke/EverythingLM-13B-16K-GPTQ",
+    "quantumaikr/open_llama_7b_hf",
+    "TheBloke/chronos-wizardlm-uc-scot-st-13B-GPTQ",
+    "TheBloke/WizardLM-30B-Uncensored-GPTQ",
+    "IDEA-CCNL/Ziya-LLaMA-13B-v1",
+    "Phind/Phind-CodeLlama-34B-v1",
+    "robowaifudev/megatron-gpt2-345m",
+    "MayaPH/GodziLLa-30B-instruct",
+    "TheBloke/CAMEL-33B-Combined-Data-SuperHOT-8K-fp16",
+    "uukuguy/speechless-codellama-orca-platypus-13b-0.10e",
+    "doas/test2",
+    "BreadAi/PM_modelV2",
+    "bigcode/santacoder",
+    "TheBloke/wizard-vicuna-13B-GPTQ",
+    "porkorbeef/Llama-2-13b",
+    "TehVenom/DiffMerge-DollyGPT-Pygmalion",
+    "PygmalionAI/pygmalion-350m",
+    "TheBloke/orca_mini_v3_7B-GPTQ",
+    "TheBloke/WizardLM-Uncensored-SuperCOT-StoryTelling-30B-GPTQ",
+    "TheBloke/WizardLM-30B-GPTQ",
+    "bigscience/bloom-560m",
+    "TFLai/gpt2-turkish-uncased",
+    "TheBloke/guanaco-33B-GPTQ",
+    "TheBloke/openchat_v2_openorca_preview-GPTQ",
+    "porkorbeef/Llama-2-13b-public",
+    "TheBloke/LongChat-13B-GPTQ",
+    "yhyhy3/med-orca-instruct-33b",
+    "TheBloke/airoboros-33B-gpt4-1-4-SuperHOT-8K-fp16",
+    "TheBloke/Chinese-Alpaca-33B-SuperHOT-8K-fp16",
+    "MayaPH/FinOPT-Franklin",
+    "TheBloke/WizardLM-33B-V1.0-Uncensored-GPTQ",
+    "TheBloke/Project-Baize-v2-13B-GPTQ",
+    "malhajar/Platypus2-70B-instruct-4bit-gptq",
+    "KoboldAI/OPT-350M-Erebus",
+    "rishiraj/bloom-560m-guanaco",
+    "Panchovix/WizardLM-33B-V1.0-Uncensored-SuperHOT-8k",
+    "doas/test5",
+    "vicgalle/alpaca-7b",
+    "beomi/KoAlpaca-Polyglot-5.8B",
+    "Phind/Phind-CodeLlama-34B-Python-v1",
+    "timdettmers/guanaco-65b-merged",
+    "TheBloke/wizard-mega-13B-GPTQ",
+    "MayaPH/GodziLLa-30B-plus",
+    "TheBloke/Platypus-30B-SuperHOT-8K-fp16",
+    "facebook/opt-350m",
+    "KoboldAI/OPT-350M-Nerys-v2",
+    "TheBloke/robin-33B-v2-GPTQ",
+    "jaspercatapang/Echidna-30B",
+    "TheBloke/llama-30b-supercot-SuperHOT-8K-fp16",
+    "marcchew/test1",
+    "Harshvir/LaMini-Neo-1.3B-Mental-Health_lora",
+    "golaxy/gogpt-560m",
+    "TheBloke/orca_mini_13B-GPTQ",
+    "Panchovix/airoboros-33b-gpt4-1.2-SuperHOT-8k",
+    "Aspik101/tulu-7b-instruct-pl-lora_unload",
+    "Phind/Phind-CodeLlama-34B-v2",
+    "BreadAi/MusePy-1-2",
+    "cerebras/Cerebras-GPT-590M",
+    "microsoft/CodeGPT-small-py",
+    "victor123/WizardLM-13B-1.0",
+    "OptimalScale/robin-65b-v2-delta",
+    "voidful/changpt-bart",
+    "FabbriSimo01/GPT_Large_Quantized",
+    "MayaPH/FinOPT-Lincoln",
+    "KoboldAI/fairseq-dense-125M",
+    "SebastianSchramm/Cerebras-GPT-111M-instruction",
+    "TheTravellingEngineer/bloom-560m-RLHF",
+    "breadlicker45/dough-instruct-base-001",
+    "WizardLM/WizardLM-30B-V1.0",
+    "WizardLM/WizardLM-30B-V1.0",
+    "WizardLM/WizardLM-30B-V1.0",
+    "TaylorAI/Flash-Llama-30M-20001",
+    "porkorbeef/Llama-2-13b-12_153950",
+    "huggingtweets/bladeecity-jerma985",
+    "KnutJaegersberg/megatron-GPT-2-345m-EvolInstruct",
+    "bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16",
+    "microsoft/DialoGPT-small",
+    "Corianas/590m",
+    "facebook/xglm-564M",
+    "EleutherAI/gpt-neo-125m",
+    "EleutherAI/pythia-160m-deduped",
+    "klosax/pythia-160m-deduped-step92k-193bt",
+    "MBZUAI/lamini-neo-125m",
+    "bigcode/tiny_starcoder_py",
+    "concedo/OPT-19M-ChatSalad",
+    "anton-l/gpt-j-tiny-random",
+    "grantprice/Cerebras-GPT-590M-finetuned-DND",
+    "deepnight-research/zsc-text",
+    "WangZeJun/bloom-820m-chat",
+    "cerebras/Cerebras-GPT-256M",
+    "ai-forever/rugpt3large_based_on_gpt2",
+    "alibidaran/medical_transcription_generator",
+    "Deci/DeciCoder-1b",
+    "microsoft/DialoGPT-medium",
+    "ogimgio/gpt-neo-125m-neurallinguisticpioneers",
+    "open-llm-leaderboard/bloom-560m-4bit-alpaca-auto-eval-adapter-applied",
+    "BreadAi/gpt-YA-1-1_160M",
+    "microsoft/DialoGPT-large",
+    "facebook/opt-125m",
+    "huggingtweets/jerma985",
+    "Locutusque/gpt2-conversational-or-qa",
+    "concedo/Pythia-70M-ChatSalad",
+    "roneneldan/TinyStories-1M",
+    "BreadAi/DiscordPy",
+    "bigcode/gpt_bigcode-santacoder",
+    "Tincando/fiction_story_generator",
+    "klosax/pythia-70m-deduped-step44k-92bt",
+    "Quake24/easyTermsSummerizer",
+    "BreadAi/gpt-YA-1-1_70M",
+    "EleutherAI/pythia-160m",
+    "euclaise/gpt-neox-122m-minipile-digits",
+    "MBZUAI/lamini-cerebras-590m",
+    "nicholasKluge/Aira-124M",
+    "MayaPH/FinOPT-Washington",
+    "cyberagent/open-calm-large",
+    "BreadAi/StoryPy",
+    "EleutherAI/pythia-70m",
+    "BreadAi/gpt-Youtube",
+    "roneneldan/TinyStories-33M",
+    "EleutherAI/pythia-70m-deduped",
+    "lgaalves/gpt2_guanaco-dolly-platypus",
+    "Corianas/Quokka_590m",
+    "lgaalves/gpt2_platypus-dolly-guanaco",
+    "cyberagent/open-calm-7b",
+    "RWKV/rwkv-4-169m-pile",
+    "gpt2",
+    "roneneldan/TinyStories-28M",
+    "lgaalves/gpt2_open-platypus",
+    "gpt2",
+    "SaylorTwift/gpt2_test",
+    "roneneldan/TinyStories-3M",
+    "nthngdy/pythia-owt2-70m-50k",
+    "Corianas/256_5epoch",
+    "roneneldan/TinyStories-8M",
+    "lgaalves/gpt2-dolly",
+    "nthngdy/pythia-owt2-70m-100k",
+    "aisquared/dlite-v2-124m",
+    "mncai/SGPT-1.3B-insurance-epoch10",
+    "huggingtweets/gladosystem",
+    "abhiramtirumala/DialoGPT-sarcastic-medium",
+    "MBZUAI/lamini-cerebras-256m",
+    "cerebras/Cerebras-GPT-111M",
+    "uberkie/metharme-1.3b-finetuned",
+    "MBZUAI/lamini-cerebras-111m",
+    "psyche/kogpt",
+    "Corianas/Quokka_256m",
+    "vicgalle/gpt2-alpaca-gpt4",
+    "aisquared/dlite-v1-124m",
+    "Mikivis/xuanxuan",
+    "MBZUAI/LaMini-GPT-124M",
+    "vicgalle/gpt2-alpaca",
+    "huashiyiqike/testmodel",
+    "Corianas/111m",
+    "baseline",
+]

src/tools/plots.py

@@ -0,0 +1,158 @@
+import numpy as np
+import pandas as pd
+import plotly.express as px
+from plotly.graph_objs import Figure
+
+from src.display.utils import  AutoEvalColumn, Task, Tasks
+from src.display.utils import human_baseline_row as HUMAN_BASELINE
+from src.leaderboard.filter_models import FLAGGED_MODELS
+from src.leaderboard.read_evals import EvalResult
+
+
+def create_scores_df(raw_data: list[EvalResult]) -> pd.DataFrame:
+    """
+    Generates a DataFrame containing the maximum scores until each date.
+
+    :param results_df: A DataFrame containing result information including metric scores and dates.
+    :return: A new DataFrame containing the maximum scores until each date for every metric.
+    """
+    # Step 1: Ensure 'date' is in datetime format and sort the DataFrame by it
+    results_df = pd.DataFrame(raw_data)
+    # results_df["date"] = pd.to_datetime(results_df["date"], format="mixed", utc=True)
+    results_df.sort_values(by="date", inplace=True)
+
+    # Step 2: Initialize the scores dictionary
+    scores = {k: [] for k in BENCHMARK_COLS + [AutoEvalColumn.average.name]}
+
+    # Step 3: Iterate over the rows of the DataFrame and update the scores dictionary
+    for task in [t.value for t in Tasks] + [Task("Average", "avg", AutoEvalColumn.average.name)]:
+        current_max = 0
+        last_date = ""
+        column = task.col_name
+        for _, row in results_df.iterrows():
+            current_model = row["full_model"]
+            # We ignore models that are flagged/no longer on the hub/not finished
+            to_ignore = (
+                not row["still_on_hub"]
+                or not row["not_flagged"]
+                or current_model in FLAGGED_MODELS
+                or row["status"] != "FINISHED"
+            )
+            if to_ignore:
+                continue
+
+            current_date = row["date"]
+            if task.benchmark == "Average":
+                current_score = np.mean(list(row["results"].values()))
+            else:
+                current_score = row["results"][task.benchmark]
+
+            if current_score > current_max:
+                if current_date == last_date and len(scores[column]) > 0:
+                    scores[column][-1] = {"model": current_model, "date": current_date, "score": current_score}
+                else:
+                    scores[column].append({"model": current_model, "date": current_date, "score": current_score})
+                current_max = current_score
+                last_date = current_date
+
+    # Step 4: Return all dictionaries as DataFrames
+    return {k: pd.DataFrame(v) for k, v in scores.items()}
+
+
+def create_plot_df(scores_df: dict[str : pd.DataFrame]) -> pd.DataFrame:
+    """
+    Transforms the scores DataFrame into a new format suitable for plotting.
+
+    :param scores_df: A DataFrame containing metric scores and dates.
+    :return: A new DataFrame reshaped for plotting purposes.
+    """
+    # Initialize the list to store DataFrames
+    dfs = []
+    # Iterate over the cols and create a new DataFrame for each column
+    for col in BENCHMARK_COLS + [AutoEvalColumn.average.name]:
+        d = scores_df[col].reset_index(drop=True)
+        d["task"] = col
+        dfs.append(d)
+
+    # Concatenate all the created DataFrames
+    concat_df = pd.concat(dfs, ignore_index=True)
+
+    # Sort values by 'date'
+    concat_df.sort_values(by="date", inplace=True)
+    concat_df.reset_index(drop=True, inplace=True)
+    return concat_df
+
+
+def create_metric_plot_obj(df: pd.DataFrame, metrics: list[str], title: str) -> Figure:
+    """
+    Create a Plotly figure object with lines representing different metrics
+    and horizontal dotted lines representing human baselines.
+
+    :param df: The DataFrame containing the metric values, names, and dates.
+    :param metrics: A list of strings representing the names of the metrics
+                    to be included in the plot.
+    :param title: A string representing the title of the plot.
+    :return: A Plotly figure object with lines representing metrics and
+             horizontal dotted lines representing human baselines.
+    """
+
+    # Filter the DataFrame based on the specified metrics
+    df = df[df["task"].isin(metrics)]
+
+    # Filter the human baselines based on the specified metrics
+    filtered_human_baselines = {k: v for k, v in HUMAN_BASELINE.items() if k in metrics}
+
+    # Create a line figure using plotly express with specified markers and custom data
+    fig = px.line(
+        df,
+        x="date",
+        y="score",
+        color="task",
+        markers=True,
+        custom_data=["task", "score", "model"],
+        title=title,
+    )
+
+    # Update hovertemplate for better hover interaction experience
+    fig.update_traces(
+        hovertemplate="<br>".join(
+            [
+                "Model Name: %{customdata[2]}",
+                "Metric Name: %{customdata[0]}",
+                "Date: %{x}",
+                "Metric Value: %{y}",
+            ]
+        )
+    )
+
+    # Update the range of the y-axis
+    fig.update_layout(yaxis_range=[0, 100])
+
+    # Create a dictionary to hold the color mapping for each metric
+    metric_color_mapping = {}
+
+    # Map each metric name to its color in the figure
+    for trace in fig.data:
+        metric_color_mapping[trace.name] = trace.line.color
+
+    # Iterate over filtered human baselines and add horizontal lines to the figure
+    for metric, value in filtered_human_baselines.items():
+        color = metric_color_mapping.get(metric, "blue")  # Retrieve color from mapping; default to blue if not found
+        location = "top left" if metric == "HellaSwag" else "bottom left"  # Set annotation position
+        # Add horizontal line with matched color and positioned annotation
+        fig.add_hline(
+            y=value,
+            line_dash="dot",
+            annotation_text=f"{metric} human baseline",
+            annotation_position=location,
+            annotation_font_size=10,
+            annotation_font_color=color,
+            line_color=color,
+        )
+
+    return fig
+
+
+# Example Usage:
+# human_baselines dictionary is defined.
+# chart = create_metric_plot_obj(scores_df, ["ARC", "HellaSwag", "MMLU", "TruthfulQA"], human_baselines, "Graph Title")

update_dynamic.py

@@ -0,0 +1,4 @@
+from src.scripts.update_all_request_files import update_dynamic_files
+
+if __name__ == "__main__":
+    update_dynamic_files()