reorg to simplify nav in code base

app.py

@@ -17,16 +17,17 @@ from src.assets.text_content import (
     LLM_BENCHMARKS_TEXT,
     TITLE,
 )
-from src.display_models.plot_results import (
+from src.plots.plot_results import (
     create_metric_plot_obj,
     create_scores_df,
     create_plot_df,
     join_model_info_with_results,
     HUMAN_BASELINES,
 )
-from src.display_models.get_model_metadata import DO_NOT_SUBMIT_MODELS, ModelType
-from src.display_models.modelcard_filter import check_model_card
-from src.display_models.utils import (
+from src.get_model_info.apply_metadata_to_df import DO_NOT_SUBMIT_MODELS, ModelType
+from src.get_model_info.get_metadata_from_hub import get_model_size
+from src.filters import check_model_card
+from src.get_model_info.utils import (
     AutoEvalColumn,
     EvalQueueColumn,
     fields,
@@ -35,8 +36,8 @@ from src.display_models.utils import (
     styled_warning,
 )
 from src.manage_collections import update_collections
-from src.load_from_hub import get_all_requested_models, get_evaluation_queue_df, get_leaderboard_df, is_model_on_hub
-from src.rate_limiting import user_submission_permission
+from src.load_from_hub import get_all_requested_models, get_evaluation_queue_df, get_leaderboard_df
+from src.filters import is_model_on_hub, user_submission_permission
 
 pd.set_option("display.precision", 1)
 
@@ -127,14 +128,8 @@ def add_new_eval(
         return styled_error("Please select a model type.")
 
     # Is the user rate limited?
-    num_models_submitted_in_period = user_submission_permission(model, users_to_submission_dates, RATE_LIMIT_PERIOD)
-    if num_models_submitted_in_period > RATE_LIMIT_QUOTA:
-        error_msg = f"Organisation or user `{model.split('/')[0]}`"
-        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 🤗"
-        )
+    user_can_submit, error_msg = user_submission_permission(model, 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?
@@ -155,28 +150,19 @@ def add_new_eval(
         if not model_on_hub:
             return styled_error(f'Model "{model}" {error}')
 
-    model_info = api.model_info(repo_id=model, revision=revision)
-
-    size_pattern = size_pattern = re.compile(r"(\d\.)?\d+(b|m)")
     try:
-        model_size = round(model_info.safetensors["total"] / 1e9, 3)
-    except AttributeError:
-        try:
-            size_match = re.search(size_pattern, model.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 65
-
-    size_factor = 8 if (precision == "GPTQ" or "GPTQ" in model) else 1
-    model_size = size_factor * model_size
+        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:
-        license = "?"
+        return styled_error("Please select a license for your model")
 
-    # Were the model card and license filled?
     modelcard_OK, error_msg = check_model_card(model)
     if not modelcard_OK:
         return styled_error(error_msg)
@@ -279,13 +265,13 @@ def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
 
 NUMERIC_INTERVALS = {
     "?": pd.Interval(-1, 0, closed="right"),
-    "0~1.5": pd.Interval(0, 1.5, closed="right"),
-    "1.5~3": pd.Interval(1.5, 3, closed="right"),
-    "3~7": pd.Interval(3, 7, closed="right"),
-    "7~13": pd.Interval(7, 13, closed="right"),
-    "13~35": pd.Interval(13, 35, closed="right"),
-    "35~60": pd.Interval(35, 60, closed="right"),
-    "60+": pd.Interval(60, 10000, 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/assets/hardcoded_evals.py

@@ -1,4 +1,4 @@
-from src.display_models.utils import AutoEvalColumn, model_hyperlink
+from src.get_model_info.utils import AutoEvalColumn, model_hyperlink
 
 gpt4_values = {
     AutoEvalColumn.model.name: model_hyperlink("https://arxiv.org/abs/2303.08774", "gpt4"),

src/assets/text_content.py

@@ -1,4 +1,4 @@
-from src.display_models.model_metadata_type import ModelType
+from src.get_model_info.hardocded_metadata.types import ModelType
 
 TITLE = """<h1 align="center" id="space-title">🤗 Open LLM Leaderboard</h1>"""
 
@@ -14,13 +14,14 @@ LLM_BENCHMARKS_TEXT = f"""
 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.
 
 ## Icons
-{ModelType.PT.to_str(" : ")} model
-{ModelType.FT.to_str(" : ")} model
-{ModelType.IFT.to_str(" : ")} model
-{ModelType.RL.to_str(" : ")} model
+{ModelType.PT.to_str(" : ")} model: new, base models, trained on a given corpora
+{ModelType.FT.to_str(" : ")} model: pretrained models finetuned on more data
+Specific fine-tune subcategories (more adapted to chat):
+{ModelType.IFT.to_str(" : ")} model: instruction fine-tunes, which are model fine-tuned specifically on datasets of task instruction 
+{ModelType.RL.to_str(" : ")} model: reinforcement fine-tunes, which usually change the model loss a bit with an added policy. 
 If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
 
-🏴‍☠️ indicates that this model has been flagged by the community, and should probably be ignored! Clicking the icon will redirect you to the discussion about the model.
+"Flagged" indicates that this model has been flagged by the community, and should probably be ignored! Clicking the link will redirect you to the discussion about the model.
 (For ex, the model was trained on the evaluation data, and is therefore cheating on the leaderboard.)
 
 ## How it works

src/{display_models/modelcard_filter.py → filters.py}

@@ -1,5 +1,8 @@
 import huggingface_hub
 from huggingface_hub import ModelCard
+from transformers import AutoConfig
+
+from datetime import datetime, timedelta, timezone
 
 
 # ht to @Wauplin, thank you for the snippet!
@@ -24,3 +27,40 @@ def check_model_card(repo_id: str) -> tuple[bool, str]:
         return False, "Please add a description to your model card, it is too short."
 
     return True, ""
+
+
+def is_model_on_hub(model_name: str, revision: str) -> bool:
+    try:
+        AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=False)
+        return True, None
+
+    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.",
+        )
+
+    except Exception:
+        return False, "was not found on hub!"
+
+
+def user_submission_permission(submission_name, users_to_submission_dates, rate_limit_period, rate_limit_quota):
+    org_or_user, _ = submission_name.split("/")
+    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 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, ""
+

src/{display_models/get_model_metadata.py → get_model_info/apply_metadata_to_df.py}

@@ -6,9 +6,9 @@ from typing import List
 from huggingface_hub import HfApi
 from tqdm import tqdm
 
-from src.display_models.model_metadata_flags import DO_NOT_SUBMIT_MODELS, FLAGGED_MODELS
-from src.display_models.model_metadata_type import MODEL_TYPE_METADATA, ModelType, model_type_from_str
-from src.display_models.utils import AutoEvalColumn, model_hyperlink
+from src.get_model_info.hardocded_metadata.flags import DO_NOT_SUBMIT_MODELS, FLAGGED_MODELS
+from src.get_model_info.hardocded_metadata.types import MODEL_TYPE_METADATA, ModelType, model_type_from_str
+from src.get_model_info.utils import AutoEvalColumn, model_hyperlink
 
 api = HfApi(token=os.environ.get("H4_TOKEN", None))
 
@@ -41,7 +41,7 @@ def get_model_metadata(leaderboard_data: List[dict]):
                 request = json.load(f)
             model_type = model_type_from_str(request["model_type"])
             model_data[AutoEvalColumn.model_type.name] = model_type.value.name
-            model_data[AutoEvalColumn.model_type_symbol.name] = model_type.value.symbol  # + ("🔺" if is_delta else "")
+            model_data[AutoEvalColumn.model_type_symbol.name] = model_type.value.symbol
             model_data[AutoEvalColumn.license.name] = request["license"]
             model_data[AutoEvalColumn.likes.name] = request["likes"]
             model_data[AutoEvalColumn.params.name] = request["params"]

src/get_model_info/get_metadata_from_hub.py

@@ -0,0 +1,19 @@
+import re
+from huggingface_hub.hf_api import ModelInfo
+
+
+def get_model_size(model_info: ModelInfo, precision: str):
+    size_pattern = size_pattern = re.compile(r"(\d\.)?\d+(b|m)")
+    try:
+        model_size = round(model_info.safetensors["total"] / 1e9, 3)
+    except AttributeError:
+        try:
+            size_match = re.search(size_pattern, model_info.modelId.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.modelId.lower()) else 1
+    model_size = size_factor * model_size
+    return model_size

src/load_from_hub.py

@@ -3,12 +3,11 @@ import os
 from collections import defaultdict
 
 import pandas as pd
-from transformers import AutoConfig
 
 from src.assets.hardcoded_evals import baseline, gpt4_values, gpt35_values
-from src.display_models.get_model_metadata import apply_metadata
-from src.display_models.read_results import get_eval_results_dicts, make_clickable_model
-from src.display_models.utils import AutoEvalColumn, EvalQueueColumn, has_no_nan_values
+from src.get_model_info.apply_metadata_to_df import apply_metadata
+from src.plots.read_results import get_eval_results_dicts, make_clickable_model
+from src.get_model_info.utils import AutoEvalColumn, EvalQueueColumn, has_no_nan_values
 
 IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
 
@@ -90,17 +89,3 @@ def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
     df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
     return df_finished[cols], df_running[cols], df_pending[cols]
 
-
-def is_model_on_hub(model_name: str, revision: str) -> bool:
-    try:
-        AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=False)
-        return True, None
-
-    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.",
-        )
-
-    except Exception:
-        return False, "was not found on hub!"

src/manage_collections.py

@@ -5,8 +5,8 @@ from requests.exceptions import HTTPError
 from huggingface_hub import get_collection, add_collection_item, update_collection_item, delete_collection_item
 from huggingface_hub.utils._errors import HfHubHTTPError
 
-from src.display_models.model_metadata_type import ModelType
-from src.display_models.utils import AutoEvalColumn
+from src.get_model_info.hardocded_metadata.types import ModelType
+from src.get_model_info.utils import AutoEvalColumn
 
 H4_TOKEN = os.environ.get("H4_TOKEN", None)
 

src/{display_models → plots}/plot_results.py

@@ -4,7 +4,7 @@ from plotly.graph_objs import Figure
 import pickle
 from datetime import datetime, timezone
 from typing import List, Dict, Tuple, Any
-from src.display_models.model_metadata_flags import FLAGGED_MODELS
+from src.get_model_info.hardocded_metadata.flags import FLAGGED_MODELS
 
 # Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
 # ARC human baseline is 0.80 (source: https://lab42.global/arc/)

src/{display_models → plots}/read_results.py

@@ -6,7 +6,7 @@ from typing import Dict, List, Tuple
 import dateutil
 import numpy as np
 
-from src.display_models.utils import AutoEvalColumn, make_clickable_model
+from src.get_model_info.utils import AutoEvalColumn, make_clickable_model
 
 METRICS = ["acc_norm", "acc_norm", "acc", "mc2"]
 BENCHMARKS = ["arc:challenge", "hellaswag", "hendrycksTest", "truthfulqa:mc"]
@@ -31,7 +31,7 @@ class EvalResult:
     date: str = ""
 
     def to_dict(self):
-        from src.load_from_hub import is_model_on_hub
+        from src.filters import is_model_on_hub
 
         if self.org is not None:
             base_model = f"{self.org}/{self.model}"

src/rate_limiting.py

@@ -1,13 +0,0 @@
-from datetime import datetime, timedelta, timezone
-
-
-def user_submission_permission(submission_name, users_to_submission_dates, rate_limit_period):
-    org_or_user, _ = submission_name.split("/")
-    if org_or_user not in users_to_submission_dates:
-        return 0
-    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]
-
-    return len(submissions_after_timelimit)