Updated model info and submit instruction

results.csv

@@ -1,16 +1,15 @@
-Models,Model Size(B),Data Source,Overall,Classification,VQA,Retrieval,Grounding
-clip-vit-large-patch14,0.428,TIGER-Lab,37.8,42.8,9.1,53.0,51.8
-blip2-opt-2.7b,3.74,TIGER-Lab,25.2,27.0,4.2,33.9,47.0
-siglip-base-patch16-224,0.203,TIGER-Lab,34.8,40.3,8.4,31.6,59.5
-open_clip-ViT-L/14,0.428,TIGER-Lab,39.7,47.8,10.9,52.3,53.3
-UniIR (BLIP_FF),0.247,TIGER-Lab,42.8,42.1,15.0,60.1,62.2
-UniIR (CLIP_SF),0.428,TIGER-Lab,44.7,44.3,16.2,61.8,65.3
-e5-v,8.36,TIGER-Lab,13.3,21.8,4.9,11.5,19.0
-Magiclens,0.428,TIGER-Lab,27.8,38.8,8.3,35.4,26.0
-CLIP-FT,0.428,TIGER-Lab,45.4,55.2,19.7,53.2,62.2
-OpenCLIP-FT,0.428,TIGER-Lab,47.2,56.0,21.9,55.4,64.1
-VLM2Vec (Phi-3.5-V-FT),4.15,TIGER-Lab,55.9,52.8,50.3,57.8,72.3
-VLM2Vec (Phi-3.5-V-LoRA),4.15,TIGER-Lab,60.1,54.8,54.9,62.3,79.5
-VLM2Vec (LLaVA-1.6-LoRA-LowRes),7.57,TIGER-Lab,55.0,54.7,50.3,56.2,64.0
-VLM2Vec (LLaVA-1.6-LoRA-HighRes),7.57,TIGER-Lab,62.9,61.2,49.9,67.4,86.1
-MMRet-large,0.428,Self-Reported,64.1,56.0,57.4,69.9,83.6
+Models,Model Size(B),Data Source,Overall,IND,OOD,Classification,VQA,Retrieval,Grounding
+CLIP,unk,unk,37.8,37.1,38.7,42.8,9.1,53.0,51.8
+BLIP2,unk,unk,25.2,25.3,25.1,27.0,4.2,33.9,47.0
+SigLIP,unk,unk,34.8,32.3,38.0,40.3,8.4,31.6,59.5
+OpenCLIP,unk,unk,39.7,39.3,40.2,47.8,10.9,52.3,53.3
+UniIR (BLIP_FF),unk,unk,42.8,44.7,40.4,42.1,15.0,60.1,62.2
+UniIR (CLIP_SF),unk,unk,44.7,47.1,41.7,44.3,16.2,61.8,65.3
+E5-V,unk,unk,13.3,14.9,11.5,21.8,4.9,11.5,19.0
+Magiclens,unk,unk,27.8,31.0,23.7,38.8,8.3,35.4,26.0
+CLIP-FFT,unk,TIGER-Lab,45.4,47.6,42.8,55.2,19.7,53.2,62.2
+OpenCLIP-FFT,unk,unk,47.2,50.5,43.1,56.0,21.9,55.4,64.1
+VLM2Vec (Phi-3.5-V-FFT),unk,TIGER-Lab,55.9,62.8,47.4,52.8,50.3,57.8,72.3
+VLM2Vec (Phi-3.5-V-LoRA),unk,TIGER-Lab,60.1,66.5,52.0,54.8,54.9,62.3,79.5
+VLM2Vec (LLaVA-1.6-LoRA-LowRes),unk,TIGER-Lab,55.0,61.0,47.5,54.7,50.3,56.2,64.0
+VLM2Vec (LLaVA-1.6-LoRA-HighRes),unk,TIGER-Lab,62.9,67.5,57.1,61.2,49.9,67.4,86.1

src/about.py

@@ -0,0 +1,72 @@
+from dataclasses import dataclass
+from enum import Enum
+
+@dataclass
+class Task:
+    benchmark: str
+    metric: str
+    col_name: str
+
+
+# Select your tasks here
+# ---------------------------------------------------
+class Tasks(Enum):
+    # task_key in the json file, metric_key in the json file, name to display in the leaderboard 
+    task0 = Task("anli_r1", "acc", "ANLI")
+    task1 = Task("logiqa", "acc_norm", "LogiQA")
+
+NUM_FEWSHOT = 0 # Change with your few shot
+# ---------------------------------------------------
+
+
+
+# Your leaderboard name
+TITLE = """<h1 align="center" id="space-title">MMEB Leaderboard</h1>"""
+
+# What does your leaderboard evaluate?
+INTRODUCTION_TEXT = """
+Intro text
+"""
+
+# Which evaluations are you running? how can people reproduce what you have?
+LLM_BENCHMARKS_TEXT = f"""
+## How it works
+
+## Reproducibility
+To reproduce our results, here is the commands you can run:
+
+"""
+
+EVALUATION_QUEUE_TEXT = """
+## Some good practices 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
+
+## In case of model failure
+If your model is displayed in the `FAILED` category, its execution stopped.
+Make sure you have followed the above steps first.
+If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).
+"""
+
+CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
+CITATION_BUTTON_TEXT = r"""
+"""

src/display/css_html_js.py

@@ -0,0 +1,106 @@
+custom_css = """
+
+.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);
+}
+
+#leaderboard-table {
+    margin-top: 15px
+}
+
+#leaderboard-table-lite {
+    margin-top: 15px
+}
+
+#search-bar-table-box > div:first-child {
+    background: none;
+    border: none;
+}
+ 
+#search-bar {
+    padding: 0px;
+}
+
+/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
+#leaderboard-table td:nth-child(2),
+#leaderboard-table th:nth-child(2) {
+    max-width: 400px;
+    overflow: auto;
+    white-space: nowrap;
+}
+
+.tab-buttons button {
+    font-size: 20px;
+    height: 1500px;
+}
+
+#scale-logo {
+    border-style: none !important;
+    box-shadow: none;
+    display: block;
+    margin-left: auto;
+    margin-right: auto;
+    max-width: 600px;
+}
+
+#scale-logo .download {
+    display: none;
+}
+#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 +1,27 @@
+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}"
+    return model_hyperlink(link, model_name)
+
+
+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 +1,110 @@
+from dataclasses import dataclass, make_dataclass
+from enum import Enum
+
+import pandas as pd
+
+from src.about import Tasks
+
+def fields(raw_class):
+    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
+
+
+# These classes are for user facing column names,
+# to avoid having to change them all around the code
+# when a modif is needed
+@dataclass
+class ColumnContent:
+    name: str
+    type: str
+    displayed_by_default: bool
+    hidden: bool = False
+    never_hidden: bool = False
+
+## Leaderboard columns
+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(["average", ColumnContent, ColumnContent("Average ⬆️", "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(["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)])
+auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
+
+# We use make dataclass to dynamically fill the scores from Tasks
+AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
+
+## For the queue columns in the submission tab
+@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)
+
+## All the model information that we might need
+@dataclass
+class ModelDetails:
+    name: str
+    display_name: str = ""
+    symbol: str = "" # emoji
+
+
+class ModelType(Enum):
+    PT = ModelDetails(name="pretrained", symbol="🟢")
+    FT = ModelDetails(name="fine-tuned", symbol="🔶")
+    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
+    RL = ModelDetails(name="RL-tuned", 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 "pretrained" in type or "🟢" in type:
+            return ModelType.PT
+        if "RL-tuned" in type or "🟦" in type:
+            return ModelType.RL
+        if "instruction-tuned" in type or "⭕" in type:
+            return ModelType.IFT
+        return ModelType.Unknown
+
+class WeightType(Enum):
+    Adapter = ModelDetails("Adapter")
+    Original = ModelDetails("Original")
+    Delta = ModelDetails("Delta")
+
+class Precision(Enum):
+    float16 = ModelDetails("float16")
+    bfloat16 = ModelDetails("bfloat16")
+    Unknown = ModelDetails("?")
+
+    def from_str(precision):
+        if precision in ["torch.float16", "float16"]:
+            return Precision.float16
+        if precision in ["torch.bfloat16", "bfloat16"]:
+            return Precision.bfloat16
+        return Precision.Unknown
+
+# Column selection
+COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
+
+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]
+

src/envs.py

@@ -0,0 +1,25 @@
+import os
+
+from huggingface_hub import HfApi
+
+# Info to change for your repository
+# ----------------------------------
+TOKEN = os.environ.get("HF_TOKEN") # A read/write token for your org
+
+OWNER = "demo-leaderboard-backend" # Change to your org - don't forget to create a results and request dataset, with the correct format!
+# ----------------------------------
+
+REPO_ID = f"{OWNER}/leaderboard"
+QUEUE_REPO = f"{OWNER}/requests"
+RESULTS_REPO = f"{OWNER}/results"
+
+# If you setup a cache later, just change HF_HOME
+CACHE_PATH=os.getenv("HF_HOME", ".")
+
+# Local caches
+EVAL_REQUESTS_PATH = os.path.join(CACHE_PATH, "eval-queue")
+EVAL_RESULTS_PATH = os.path.join(CACHE_PATH, "eval-results")
+EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")
+EVAL_RESULTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-results-bk")
+
+API = HfApi(token=TOKEN)

src/leaderboard/read_evals.py

@@ -0,0 +1,196 @@
+import glob
+import json
+import math
+import os
+from dataclasses import dataclass
+
+import dateutil
+import numpy as np
+
+from src.display.formatting import make_clickable_model
+from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
+from src.submission.check_validity import is_model_on_hub
+
+
+@dataclass
+class EvalResult:
+    """Represents one full evaluation. Built from a combination of the result and request file for a given run.
+    """
+    eval_name: str # org_model_precision (uid)
+    full_model: str # org/model (path on hub)
+    org: str 
+    model: str
+    revision: str # commit hash, "" if main
+    results: dict
+    precision: Precision = Precision.Unknown
+    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
+    weight_type: WeightType = WeightType.Original # Original or Adapter
+    architecture: str = "Unknown" 
+    license: str = "?"
+    likes: int = 0
+    num_params: int = 0
+    date: str = "" # submission date of request file
+    still_on_hub: bool = False
+
+    @classmethod
+    def init_from_json_file(self, json_filepath):
+        """Inits the result from the specific model result file"""
+        with open(json_filepath) as fp:
+            data = json.load(fp)
+
+        config = data.get("config")
+
+        # Precision
+        precision = Precision.from_str(config.get("model_dtype"))
+
+        # Get model and org
+        org_and_model = config.get("model_name", config.get("model_args", None))
+        org_and_model = org_and_model.split("/", 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)
+
+        still_on_hub, _, model_config = is_model_on_hub(
+            full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
+        )
+        architecture = "?"
+        if model_config is not None:
+            architectures = getattr(model_config, "architectures", None)
+            if architectures:
+                architecture = ";".join(architectures)
+
+        # Extract results available in this file (some results are split in several files)
+        results = {}
+        for task in Tasks:
+            task = task.value
+
+            # We average all scores of a given metric (not all metrics are present in all files)
+            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == 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 self(
+            eval_name=result_key,
+            full_model=full_model,
+            org=org,
+            model=model,
+            results=results,
+            precision=precision,  
+            revision= config.get("model_sha", ""),
+            still_on_hub=still_on_hub,
+            architecture=architecture
+        )
+
+    def update_with_request_file(self, requests_path):
+        """Finds the relevant request file for the current model and updates info with it"""
+        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
+
+        try:
+            with open(request_file, "r") as f:
+                request = json.load(f)
+            self.model_type = ModelType.from_str(request.get("model_type", ""))
+            self.weight_type = WeightType[request.get("weight_type", "Original")]
+            self.license = request.get("license", "?")
+            self.likes = request.get("likes", 0)
+            self.num_params = request.get("params", 0)
+            self.date = request.get("submitted_time", "")
+        except Exception:
+            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
+
+    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.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,
+        }
+
+        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"""
+    request_files = os.path.join(
+        requests_path,
+        f"{model_name}_eval_request_*.json",
+    )
+    request_files = glob.glob(request_files)
+
+    # Select correct request file (precision)
+    request_file = ""
+    request_files = sorted(request_files, reverse=True)
+    for tmp_request_file in request_files:
+        with open(tmp_request_file, "r") as f:
+            req_content = json.load(f)
+            if (
+                req_content["status"] in ["FINISHED"]
+                and req_content["precision"] == precision.split(".")[-1]
+            ):
+                request_file = tmp_request_file
+    return request_file
+
+
+def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
+    """From the path of the results folder root, extract all needed info for results"""
+    model_result_filepaths = []
+
+    for root, _, files in os.walk(results_path):
+        # We should only have json files in model results
+        if len(files) == 0 or any([not f.endswith(".json") for f in files]):
+            continue
+
+        # Sort the files by date
+        try:
+            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
+        except dateutil.parser._parser.ParserError:
+            files = [files[-1]]
+
+        for file in files:
+            model_result_filepaths.append(os.path.join(root, file))
+
+    eval_results = {}
+    for model_result_filepath in model_result_filepaths:
+        # Creation of result
+        eval_result = EvalResult.init_from_json_file(model_result_filepath)
+        eval_result.update_with_request_file(requests_path)
+
+        # 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 v in eval_results.values():
+        try:
+            v.to_dict() # we test if the dict version is complete
+            results.append(v)
+        except KeyError:  # not all eval values present
+            continue
+
+    return results

src/populate.py

@@ -0,0 +1,58 @@
+import json
+import os
+
+import pandas as pd
+
+from src.display.formatting import has_no_nan_values, make_clickable_model
+from src.display.utils import AutoEvalColumn, EvalQueueColumn
+from src.leaderboard.read_evals import get_raw_eval_results
+
+
+def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
+    """Creates a dataframe from all the individual experiment results"""
+    raw_data = get_raw_eval_results(results_path, requests_path)
+    all_data_json = [v.to_dict() for v in raw_data]
+
+    df = pd.DataFrame.from_records(all_data_json)
+    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
+    df = df[cols].round(decimals=2)
+
+    # filter out if any of the benchmarks have not been produced
+    df = df[has_no_nan_values(df, benchmark_cols)]
+    return df
+
+
+def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
+    """Creates the different dataframes for the evaluation queues requestes"""
+    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
+    all_evals = []
+
+    for entry in entries:
+        if ".json" in entry:
+            file_path = os.path.join(save_path, entry)
+            with open(file_path) as fp:
+                data = json.load(fp)
+
+            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
+            data[EvalQueueColumn.revision.name] = data.get("revision", "main")
+
+            all_evals.append(data)
+        elif ".md" not in entry:
+            # this is a folder
+            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
+            for sub_entry in sub_entries:
+                file_path = os.path.join(save_path, entry, sub_entry)
+                with open(file_path) as fp:
+                    data = json.load(fp)
+
+                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
+                data[EvalQueueColumn.revision.name] = data.get("revision", "main")
+                all_evals.append(data)
+
+    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
+    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
+    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
+    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
+    df_running = pd.DataFrame.from_records(running_list, columns=cols)
+    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
+    return df_finished[cols], df_running[cols], df_pending[cols]

src/submission/check_validity.py

@@ -0,0 +1,99 @@
+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
+from transformers import AutoConfig
+from transformers.models.auto.tokenization_auto import AutoTokenizer
+
+def check_model_card(repo_id: str) -> tuple[bool, str]:
+    """Checks if the model card and license exist and have been filled"""
+    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."
+
+    # 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."
+            )
+
+    # Enforce card content
+    if len(card.text) < 200:
+        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, token: str = None, trust_remote_code=False, test_tokenizer=False) -> tuple[bool, str]:
+    """Checks if the model model_name is on the hub, and whether it (and its tokenizer) can be loaded with AutoClasses."""
+    try:
+        config = AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
+        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 as e:
+                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:
+        return False, "was not found on hub!", None
+
+
+def get_model_size(model_info: ModelInfo, precision: str):
+    """Gets the model size from the configuration, or the model name if the configuration does not contain the information."""
+    try:
+        model_size = round(model_info.safetensors["total"] / 1e9, 3)
+    except (AttributeError, TypeError):
+        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
+
+def get_model_arch(model_info: ModelInfo):
+    """Gets the model architecture from the configuration"""
+    return model_info.config.get("architectures", "Unknown")
+
+def already_submitted_models(requested_models_dir: str) -> set[str]:
+    """Gather a list of already submitted models to avoid duplicates"""
+    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

src/submission/submit.py

@@ -0,0 +1,119 @@
+import json
+import os
+from datetime import datetime, timezone
+
+from src.display.formatting import styled_error, styled_message, styled_warning
+from src.envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
+from src.submission.check_validity import (
+    already_submitted_models,
+    check_model_card,
+    get_model_size,
+    is_model_on_hub,
+)
+
+REQUESTED_MODELS = None
+USERS_TO_SUBMISSION_DATES = None
+
+def add_new_eval(
+    model: str,
+    base_model: str,
+    revision: str,
+    precision: str,
+    weight_type: str,
+    model_type: 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.")
+
+    # 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=TOKEN, test_tokenizer=True)
+        if not base_model_on_hub:
+            return styled_error(f'Base model "{base_model}" {error}')
+
+    if not weight_type == "Adapter":
+        model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, token=TOKEN, test_tokenizer=True)
+        if not model_on_hub:
+            return styled_error(f'Model "{model}" {error}')
+
+    # 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 = check_model_card(model)
+    if not modelcard_OK:
+        return styled_error(error_msg)
+
+    # Seems good, creating the eval
+    print("Adding new eval")
+
+    eval_entry = {
+        "model": model,
+        "base_model": base_model,
+        "revision": revision,
+        "precision": precision,
+        "weight_type": weight_type,
+        "status": "PENDING",
+        "submitted_time": current_time,
+        "model_type": model_type,
+        "likes": model_info.likes,
+        "params": model_size,
+        "license": license,
+        "private": False,
+    }
+
+    # 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_False_{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",
+    )
+
+    # 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 for the model to show in the PENDING list."
+    )

utils.py

@@ -13,14 +13,14 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
 TASKS = ["Classification", "VQA", "Retrieval", "Grounding"]
 
 MODEL_INFO = [
-    "Rank", "Models", "Model Size(B)", "Data Source",
-    "Overall",
+    "Models", "Model Size(B)", "Data Source",
+    "Overall", "IND", "OOD", 
     "Classification", "VQA", "Retrieval", "Grounding"
 ]
 
 BASE_COLS = [col for col in MODEL_INFO if col not in TASKS]
 
-DATA_TITLE_TYPE = ['number', 'markdown', 'str', 'markdown', 'number', 'number', 'number', 'number', 'number', 'number', 'number']
+DATA_TITLE_TYPE = ['markdown', 'str', 'markdown', 'number', 'number', 'number', 'number', 'number', 'number', 'number']
 
 SUBMISSION_NAME = "MMEB"
 SUBMISSION_URL = os.path.join("https://huggingface.co/spaces/TIGER-Lab/", SUBMISSION_NAME)
@@ -69,12 +69,7 @@ to the VQA task, each query has 1,000 candidates, with 1 ground truth and 999 di
 """
 
 CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
-CITATION_BUTTON_TEXT = r"""@article{jiang2024vlm2vec,
-  title={VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks},
-  author={Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu},
-  journal={arXiv preprint arXiv:2410.05160},
-  year={2024}
-}"""
+CITATION_BUTTON_TEXT = """"""
 
 SUBMIT_INTRODUCTION = """# Submit on MMEB Leaderboard Introduction
 
@@ -82,56 +77,20 @@ SUBMIT_INTRODUCTION = """# Submit on MMEB Leaderboard Introduction
 ```json
 [
     {
-        "Model": "<Model Name>",
-        <Optional>"URL": "<Model URL>",
-        "Model Size(B)": 1000,
-        "Data Source": TIGER-Lab,
-        "Overall": 50.0,
-        "Classification": 50.0,
-        "VQA": 50.0,
-        "Retrieval": 50.0,
-        "Grounding": 50.0
-    }, 
+        "question_id": 123,
+        "question": "abc",
+        "options": ["abc", "xyz", ...], 
+        "answer": "ABC",
+        "answer_index": 1,
+        "category": "abc,
+        "pred": "B",
+        "model_outputs": ""
+    }, ...
 ]
 ```
-Please send us an email at m7su@uwaterloo.ca, attaching the JSON file. We will review your submission and update the leaderboard accordingly.
+...
 """
 
-MODEL_URLS = {
-    "clip-vit-large-patch14": "https://huggingface.co/openai/clip-vit-large-patch14", 
-    "blip2-opt-2.7b": "https://huggingface.co/Salesforce/blip2-opt-2.7b",
-    "siglip-base-patch16-224": "https://huggingface.co/google/siglip-base-patch16-224", 
-    "open_clip-ViT-L/14": "https://github.com/mlfoundations/open_clip", 
-    "e5-v": "https://huggingface.co/royokong/e5-v", 
-    "Magiclens": "https://github.com/google-deepmind/magiclens",
-    "MMRet-large": "https://huggingface.co/JUNJIE99/MMRet-large", 
-    "VLM2Vec-Phi-3.5-v": "https://huggingface.co/TIGER-Lab/VLM2Vec-Full",
-    "VLM2Vec": "https://github.com/TIGER-AI-Lab/VLM2Vec", 
-    "UniIR": "https://huggingface.co/TIGER-Lab/UniIR", 
-    "OpenCLIP-FT": "https://doi.org/10.48550/arXiv.2212.07143", 
-    "CLIP-FT": "https://doi.org/10.48550/arXiv.2103.00020"
-}
-
-def create_hyperlinked_names(df):
-    def convert_url(url, model_name):
-        return f'<a href="{url}">{model_name}</a>'
-
-    def add_link_to_model_name(model_name):
-        if "VLM2Vec (Phi-3.5-V-" in model_name:
-            url = MODEL_URLS["VLM2Vec-Phi-3.5-v"]
-            return convert_url(url, model_name)
-        if "VLM2Vec (LLaVA-1.6-LoRA-" in model_name:
-            url = MODEL_URLS["VLM2Vec"]
-            return convert_url(url, model_name)
-        if "UniIR" in model_name:
-            url = MODEL_URLS["UniIR"]
-            return convert_url(url, model_name)
-        return convert_url(MODEL_URLS[model_name], model_name) if model_name in MODEL_URLS else model_name
-    
-    df = df.copy()
-    df['Models'] = df['Models'].apply(add_link_to_model_name)
-    return df
-
 def get_df():
     # fetch the leaderboard data
     url = "https://huggingface.co/spaces/TIGER-Lab/MMEB/resolve/main/results.csv"
@@ -143,8 +102,6 @@ def get_df():
     df.to_csv(CSV_DIR, index=False) # update local file
     df['Model Size(B)'] = df['Model Size(B)'].apply(process_model_size)
     df = df.sort_values(by=['Overall'], ascending=False)
-    df = create_hyperlinked_names(df)
-    df['Rank'] = range(1, len(df) + 1)
     return df
 
 
@@ -156,12 +113,10 @@ def add_new_eval(input_file):
     upload_data = json.loads(input_file)
     print("upload_data:\n", upload_data)
     data_row = [f'{upload_data["Model"]}']
-    for col in ['Overall', 'Model Size(B)'] + TASKS:
+    for col in ['Overall', 'Model Size(B)', 'IND', 'OOD'] + TASKS:
         if not col in upload_data.keys():
             return f"Error! Missing {col} column!"
         data_row += [upload_data[col]]
-    if 'URL' in upload_data.keys():
-        MODEL_URLS[upload_data['Model']] = upload_data['URL']
     print("data_row:\n", data_row)
     submission_repo = Repository(local_dir=SUBMISSION_NAME, clone_from=SUBMISSION_URL,
                                  use_auth_token=HF_TOKEN, repo_type="space")
@@ -204,11 +159,35 @@ def search_and_filter_models(df, query, min_size, max_size):
     return filtered_df[COLUMN_NAMES]
 
 
+# def search_and_filter_models(df, query, min_size, max_size):
+#     filtered_df = df.copy()
+
+#     if query:
+#         filtered_df = filtered_df[filtered_df['Models'].str.contains(query, case=False, na=False)]
+    
+#     def size_filter(x):
+#         if isinstance(x, (int, float)):
+#             return min_size <= x <= max_size
+#         return True 
+    
+#     filtered_df = filtered_df[filtered_df['Model Size(B)'].apply(size_filter)]
+    
+#     return filtered_df[COLUMN_NAMES]
+
+
 def search_models(df, query):
     if query:
         return df[df['Models'].str.contains(query, case=False, na=False)]
     return df
 
+
+# def get_size_range(df):
+#     numeric_sizes = df[df['Model Size(B)'].apply(lambda x: isinstance(x, (int, float)))]['Model Size(B)']
+#     if len(numeric_sizes) > 0:
+#         return float(numeric_sizes.min()), float(numeric_sizes.max())
+#     return 0, 1000
+
+
 def get_size_range(df):
     sizes = df['Model Size(B)'].apply(lambda x: 0.0 if x == 'unknown' else x)
     if (sizes == 0.0).all():