feat: add mvp leaderboard

app.py

@@ -1,204 +1,208 @@
+import os
+
 import gradio as gr
-from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
 import pandas as pd
 from apscheduler.schedulers.background import BackgroundScheduler
-from huggingface_hub import snapshot_download
-
-from src.about import (
-    CITATION_BUTTON_LABEL,
-    CITATION_BUTTON_TEXT,
-    EVALUATION_QUEUE_TEXT,
-    INTRODUCTION_TEXT,
-    LLM_BENCHMARKS_TEXT,
-    TITLE,
-)
-from src.display.css_html_js import custom_css
-from src.display.utils import (
-    BENCHMARK_COLS,
-    COLS,
-    EVAL_COLS,
-    EVAL_TYPES,
-    AutoEvalColumn,
-    ModelType,
-    fields,
-    WeightType,
-    Precision
+from huggingface_hub import HfApi
+
+from src.about import CITATION_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE
+from src.populate import (
+    MU_MATH_COLUMNS_DICT,
+    U_MATH_COLUMNS_DICT,
+    Field,
+    get_mu_math_leaderboard_df,
+    get_u_math_leaderboard_df,
 )
-from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
-from src.populate import get_evaluation_queue_df, get_leaderboard_df
-from src.submission.submit import add_new_eval
 
 
 def restart_space():
+    TOKEN = os.environ.get("HF_TOKEN")  # A read/write token for your org
+    API = HfApi(token=TOKEN)
+    REPO_ID = "toloka/u-math-leaderboard"
     API.restart_space(repo_id=REPO_ID)
 
-### Space initialisation
-try:
-    print(EVAL_REQUESTS_PATH)
-    snapshot_download(
-        repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-try:
-    print(EVAL_RESULTS_PATH)
-    snapshot_download(
-        repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-
-
-LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
-
-(
-    finished_eval_queue_df,
-    running_eval_queue_df,
-    pending_eval_queue_df,
-) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
-
-def init_leaderboard(dataframe):
+
+LEADERBOARD_U_MATH_DF = get_u_math_leaderboard_df()
+LEADERBOARD_MU_MATH_DF = get_mu_math_leaderboard_df()
+
+
+def init_leaderboard(dataframe: pd.DataFrame, columns_dict: dict[str, Field]) -> gr.components.Component:
     if dataframe is None or dataframe.empty:
         raise ValueError("Leaderboard DataFrame is empty or None.")
-    return Leaderboard(
-        value=dataframe,
-        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],
-            label="Select Columns to Display:",
-        ),
-        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
-        hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
-        filter_columns=[
-            ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"),
-            ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"),
-            ColumnFilter(
-                AutoEvalColumn.params.name,
-                type="slider",
-                min=0.01,
-                max=150,
-                label="Select the number of parameters (B)",
-            ),
-            ColumnFilter(
-                AutoEvalColumn.still_on_hub.name, type="boolean", label="Deleted/incomplete", default=True
-            ),
-        ],
-        bool_checkboxgroup_label="Hide models",
-        interactive=False,
-    )
-
-
-demo = gr.Blocks(css=custom_css)
+
+    def filter_dataframe_by_selected_columns(full_df: pd.DataFrame, columns: list[str]) -> pd.DataFrame:
+        always_here_cols = [c.pretty_name for c in columns_dict.values() if c.never_hidden]
+        selected_columns = [c for c in columns if c in full_df.columns and c not in always_here_cols]
+        # keep the order of the columns
+        filtered_df = full_df[[c for c in full_df.columns if c in (always_here_cols + selected_columns)]]
+        return filtered_df
+
+    def filter_dataframe_by_selected_tag_columns(
+        full_df: pd.DataFrame, current_tag: str
+    ) -> tuple[pd.DataFrame, list[str], str, str]:
+        always_here_cols = [c.pretty_name for c in columns_dict.values() if c.never_hidden]
+        selected_columns = [
+            c.pretty_name for c in columns_dict.values() if current_tag in c.tags and c not in always_here_cols
+        ]
+        # keep the order of the columns
+        filtered_df = full_df[[c for c in full_df.columns if c in (always_here_cols + selected_columns)]]
+        _columns_to_select_visibility = [
+            c.pretty_name for c in columns_dict.values() if not c.fully_hidden and not c.never_hidden
+        ]
+        return filtered_df, [c for c in _columns_to_select_visibility if c in filtered_df.columns], "All", "All"
+
+    def filter_dataframe_by_search(full_df: pd.DataFrame, current_df: pd.DataFrame, search: str) -> pd.DataFrame:
+        filtered_df = full_df[
+            (full_df[columns_dict["model_name"].pretty_name].str.contains(search, case=False, na=False))
+        ]
+        return filtered_df[current_df.columns]
+
+    def filter_dataframe_by_model_type(
+        full_df: pd.DataFrame, current_df: pd.DataFrame, filter_name: str
+    ) -> pd.DataFrame:
+        if filter_name == "All":
+            return full_df[current_df.columns]
+        else:
+            # actually filter by emoji
+            query_symbol = filter_name[0]
+            filtered_df = full_df[full_df[columns_dict["model_type_symbol"].pretty_name] == query_symbol]
+            return filtered_df[current_df.columns]
+
+    def filter_dataframe_by_model_size(
+        full_df: pd.DataFrame, current_df: pd.DataFrame, filter_name: str
+    ) -> pd.DataFrame:
+        if filter_name == "All":
+            return full_df[current_df.columns]
+        else:
+            # actually filter by emoji
+            query_symbol = filter_name[0]
+            filtered_df = full_df[full_df[columns_dict["model_size_symbol"].pretty_name] == query_symbol]
+            return filtered_df[current_df.columns]
+
+    with gr.Column() as col:
+        # Add the controls
+        with gr.Accordion("➡️ See All Columns", open=False):
+            columns_to_select_visibility = [
+                c.pretty_name for c in columns_dict.values() if not c.fully_hidden and not c.never_hidden
+            ]
+            all_columns_selector = gr.CheckboxGroup(
+                choices=columns_to_select_visibility,
+                value=[
+                    c.pretty_name
+                    for c in columns_dict.values()
+                    if c.pretty_name in columns_to_select_visibility and c.displayed_by_default
+                ],
+                label="Select Columns to Display:",
+                interactive=True,
+                container=False,
+            )
+        with gr.Row():
+            with gr.Column():
+                search_bar = gr.Textbox(
+                    placeholder="🔍 Search for your model and press ENTER...",
+                    show_label=False,
+                    elem_id="search-bar",
+                )
+
+            # collect all column tags and create buttons for them
+            all_tags = {}
+            with gr.Column(variant="panel"):
+                gr.Markdown("Select Columns:")
+                for c in columns_dict.values():
+                    for tag in c.tags:
+                        if tag not in all_tags:
+                            all_tags[tag] = gr.Button(tag, interactive=True, size="sm")
+
+            model_type_filter_selector = gr.Radio(
+                label="Filter model types:",
+                choices=["All", "💙 Open-Weights", "🟥 Proprietary"],
+                value="All",
+                elem_id="model-type-filter",
+                interactive=True,
+            )
+
+            model_size_filter_selector = gr.Radio(
+                label="Filter model sizes:",
+                choices=["All", "🛴 Tiny (<5B)", "🚗 Small (5-50B)", "🚚 Medium (50-100B)", "🚀 Large (>100B)"],
+                value="All",
+                elem_id="model-size-filter",
+                interactive=True,
+            )
+
+        # create the hidden and visible dataframes to display
+        hidden_leaderboard_df = gr.components.Dataframe(
+            value=dataframe,
+            datatype=[c.column_type for c in columns_dict.values()],
+            visible=False,
+            interactive=False,
+        )
+        leaderboard_df = gr.components.Dataframe(
+            value=dataframe[[c.pretty_name for c in columns_dict.values() if c.displayed_by_default]],
+            datatype=[c.column_type for c in columns_dict.values()],
+            elem_id="leaderboard-df",
+            interactive=False,
+        )
+        # add the callbacks
+        all_columns_selector.change(
+            fn=filter_dataframe_by_selected_columns,
+            inputs=[hidden_leaderboard_df, all_columns_selector],
+            outputs=[leaderboard_df],
+        )
+        search_bar.submit(
+            fn=filter_dataframe_by_search,
+            inputs=[hidden_leaderboard_df, leaderboard_df, search_bar],
+            outputs=[leaderboard_df],
+        )
+        model_type_filter_selector.change(
+            fn=filter_dataframe_by_model_type,
+            inputs=[hidden_leaderboard_df, leaderboard_df, model_type_filter_selector],
+            outputs=[leaderboard_df],
+        )
+        model_size_filter_selector.change(
+            fn=filter_dataframe_by_model_size,
+            inputs=[hidden_leaderboard_df, leaderboard_df, model_size_filter_selector],
+            outputs=[leaderboard_df],
+        )
+        for tag, button in all_tags.items():
+            button.click(
+                fn=filter_dataframe_by_selected_tag_columns,
+                inputs=[hidden_leaderboard_df, button],
+                outputs=[leaderboard_df, all_columns_selector, model_type_filter_selector, model_size_filter_selector],
+            )
+
+        # reload the leaderboard on the first load
+        filter_dataframe_by_selected_columns(dataframe, all_columns_selector.value)
+        return col
+
+
+demo = gr.Blocks()
 with demo:
     gr.HTML(TITLE)
     gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
 
+    print(LEADERBOARD_U_MATH_DF)
+    print(LEADERBOARD_MU_MATH_DF)
+
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
-        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
-            leaderboard = init_leaderboard(LEADERBOARD_DF)
+        with gr.TabItem("🏆 U-MATH", elem_id="u-math-benchmark-tab-table", id=0):
+            leaderboard_umath = init_leaderboard(LEADERBOARD_U_MATH_DF, U_MATH_COLUMNS_DICT)
 
-        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
-            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
+        with gr.TabItem("🏅 μ-MATH (Meta-Benchmark)", elem_id="mu-math-benchmark-tab-table", id=1):
+            leaderboard_mumath = init_leaderboard(LEADERBOARD_MU_MATH_DF, MU_MATH_COLUMNS_DICT)
 
-        with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3):
-            with gr.Column():
-                with gr.Row():
-                    gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
-
-                with gr.Column():
-                    with gr.Accordion(
-                        f"✅ Finished Evaluations ({len(finished_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            finished_eval_table = gr.components.Dataframe(
-                                value=finished_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-                    with gr.Accordion(
-                        f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            running_eval_table = gr.components.Dataframe(
-                                value=running_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-
-                    with gr.Accordion(
-                        f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            pending_eval_table = gr.components.Dataframe(
-                                value=pending_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-            with gr.Row():
-                gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")
-
-            with gr.Row():
-                with gr.Column():
-                    model_name_textbox = gr.Textbox(label="Model name")
-                    revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
-                    model_type = gr.Dropdown(
-                        choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
-                        label="Model type",
-                        multiselect=False,
-                        value=None,
-                        interactive=True,
-                    )
-
-                with gr.Column():
-                    precision = gr.Dropdown(
-                        choices=[i.value.name for i in Precision if i != Precision.Unknown],
-                        label="Precision",
-                        multiselect=False,
-                        value="float16",
-                        interactive=True,
-                    )
-                    weight_type = gr.Dropdown(
-                        choices=[i.value.name for i in WeightType],
-                        label="Weights type",
-                        multiselect=False,
-                        value="Original",
-                        interactive=True,
-                    )
-                    base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
-
-            submit_button = gr.Button("Submit Eval")
-            submission_result = gr.Markdown()
-            submit_button.click(
-                add_new_eval,
-                [
-                    model_name_textbox,
-                    base_model_name_textbox,
-                    revision_name_textbox,
-                    precision,
-                    weight_type,
-                    model_type,
-                ],
-                submission_result,
-            )
+        with gr.TabItem("📝 About", elem_id="about-tab-table", id=2):
+            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
 
-    with gr.Row():
-        with gr.Accordion("📙 Citation", open=False):
-            citation_button = gr.Textbox(
-                value=CITATION_BUTTON_TEXT,
-                label=CITATION_BUTTON_LABEL,
-                lines=20,
-                elem_id="citation-button",
-                show_copy_button=True,
-            )
+        citation_button = gr.Textbox(
+            value=CITATION_TEXT,
+            label="📙 Citation",
+            lines=9,
+            elem_id="citation-button",
+            show_copy_button=True,
+            container=True,
+        )
 
 scheduler = BackgroundScheduler()
-scheduler.add_job(restart_space, "interval", seconds=1800)
+scheduler.add_job(restart_space, "interval", seconds=60 * 60)
 scheduler.start()
-demo.queue(default_concurrency_limit=40).launch()
+demo.queue(default_concurrency_limit=40).launch()

data/mu_math_eval_results.json

@@ -0,0 +1,20 @@
+[
+  {
+    "model_name": "mistralai/Ministral-8B-Instruct-2410",
+    "extract_model_name": "Qwen/Qwen2.5-72B-Instruct",
+    "mu_math": [0.664, 0.33, 0.651, 0.68, 0.701, 0.628],
+    "GPT-4o": [0.664, 0.332, 0.621, 0.71, 0.696, 0.637],
+    "Gemini-1.5-Pro": [0.672, 0.279, 0.709, 0.585, 0.798, 0.466],
+    "Llama-3.1-70B-Instruct": [0.675, 0.317, 0.619, 0.707, 0.541, 0.769],
+    "Qwen2.5-72B-Instruct": [0.646, 0.295, 0.626, 0.672, 0.719, 0.574]
+  },
+  {
+    "model_name": "meta-llama/Llama-3.3-70B-Instruct",
+    "extract_model_name": "Qwen/Qwen2.5-72B-Instruct",
+    "mu_math": [0.741, 0.496, 0.666, 0.827, 0.816, 0.682],
+    "GPT-4o": [0.731, 0.475, 0.636, 0.832, 0.802, 0.681],
+    "Gemini-1.5-Pro": [0.705, 0.394, 0.693, 0.732, 0.856, 0.508],
+    "Llama-3.1-70B-Instruct": [0.823, 0.605, 0.67, 0.908, 0.802, 0.832],
+    "Qwen2.5-72B-Instruct": [0.705, 0.421, 0.658, 0.767, 0.791, 0.627]
+  }
+]

data/u_math_eval_results.json

@@ -0,0 +1,13 @@
+[
+  {
+    "model_name": "gpt-4o-mini-2024-07-18",
+    "judge_model_name": "gpt-4o-mini-2024-07-18",
+    "u_math": [0.5123, 0.2345, 0.1234],
+    "differential_calc": [0.5123, 0.2345, 0.1234],
+    "integral_calc": [0.43, 0.23, 0.34],
+    "algebra": [0.98, 0.12, 0.34],
+    "multivariable_calculus": [0.98, 0.12, 0.34],
+    "precalculus_review": [0.8412, 0.1234, 0.1234],
+    "sequences_series": [0.1234, 0.1234, 0.1234]
+  }
+]

pyproject.toml

@@ -1,8 +1,10 @@
 [tool.ruff]
+line-length = 119
+
+[tool.ruff.lint]
 # Enable pycodestyle (`E`) and Pyflakes (`F`) codes by default.
 select = ["E", "F"]
 ignore = ["E501"] # line too long (black is taking care of this)
-line-length = 119
 fixable = ["A", "B", "C", "D", "E", "F", "G", "I", "N", "Q", "S", "T", "W", "ANN", "ARG", "BLE", "COM", "DJ", "DTZ", "EM", "ERA", "EXE", "FBT", "ICN", "INP", "ISC", "NPY", "PD", "PGH", "PIE", "PL", "PT", "PTH", "PYI", "RET", "RSE", "RUF", "SIM", "SLF", "TCH", "TID", "TRY", "UP", "YTT"]
 
 [tool.isort]

requirements.txt

@@ -1,9 +1,9 @@
 APScheduler
 black
+isort
 datasets
 gradio
 gradio[oauth]
-gradio_leaderboard==0.0.13
 gradio_client
 huggingface-hub>=0.18.0
 matplotlib
@@ -13,4 +13,4 @@ python-dateutil
 tqdm
 transformers
 tokenizers>=0.15.0
-sentencepiece
+sentencepiece

src/about.py

@@ -1,6 +1,7 @@
 from dataclasses import dataclass
 from enum import Enum
 
+
 @dataclass
 class Task:
     benchmark: str
@@ -11,17 +12,16 @@ class Task:
 # 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 
+    # 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">Demo leaderboard</h1>"""
+TITLE = """<h1 align="center" id="space-title">U-MATH / μ-MATH leaderboard</h1>"""
 
 # What does your leaderboard evaluate?
 INTRODUCTION_TEXT = """
@@ -29,7 +29,7 @@ Intro text
 """
 
 # Which evaluations are you running? how can people reproduce what you have?
-LLM_BENCHMARKS_TEXT = f"""
+LLM_BENCHMARKS_TEXT = """
 ## How it works
 
 ## Reproducibility
@@ -37,36 +37,12 @@ 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"""
-"""
+CITATION_TEXT = r"""@misc{chernyshev2024umath,
+      title={U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs}, 
+      author={Konstantin Chernyshev and Vitaliy Polshkov and Ekaterina Artemova and Alex Myasnikov and Vlad Stepanov and Alexei Miasnikov and Sergei Tilga},
+      year={2024},
+      eprint={2412.03205},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2412.03205}, 
+}"""

src/display/css_html_js.py

@@ -1,105 +0,0 @@
-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;
-}
-
-#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

@@ -1,27 +0,0 @@
-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

@@ -1,110 +0,0 @@
-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

@@ -1,25 +0,0 @@
-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

@@ -1,196 +0,0 @@
-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

@@ -1,58 +1,286 @@
 import json
 import os
+from dataclasses import dataclass, field
 
 import pandas as pd
+from huggingface_hub import model_info
+from transformers import AutoConfig
 
-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 is_model_on_hub(
+    model_name: str, revision: str, token: str = None, trust_remote_code=False
+) -> tuple[bool, str | None, str | None]:
+    """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
+        )
+        return True, None, config
+    except Exception:
+        return False, "was not found on hub!", None
 
-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)
+def model_size_to_symbol(model_size_in_b_params: int | None) -> str:
+    """Converts model size to a symbol"""
+    if model_size_in_b_params is None or model_size_in_b_params == 0 or not model_size_in_b_params:
+        return "❓"
+    if model_size_in_b_params < 5:
+        return "🛴"
+    elif model_size_in_b_params < 50:
+        return "🚗"
+    elif model_size_in_b_params < 100:
+        return "🚚"
+    else:
+        return "🚀"
 
-    # filter out if any of the benchmarks have not been produced
-    df = df[has_no_nan_values(df, benchmark_cols)]
+
+def model_type_to_symbol(model_type: str) -> str:
+    """Converts model type to a symbol"""
+    if model_type == "Open-Weights":
+        return "💙"
+    elif model_type == "Proprietary":
+        return "🟥"
+    else:
+        return "❓"
+
+
+def get_hf_data_by_model_name(model_name: str) -> dict:
+    """Get model data from Hugging Face API by model name"""
+    still_on_hub, _, model_config = is_model_on_hub(model_name, "main", trust_remote_code=True)
+
+    architecture = "Unknown"
+    if model_config is not None:
+        architectures = getattr(model_config, "architectures", None)
+        if architectures:
+            architecture = ";".join(architectures)
+
+    num_params = None
+    if still_on_hub:
+        info = model_info(repo_id=model_name)
+        try:
+            num_params = round(info.safetensors["total"] / 1e9, 3)
+        except AttributeError as e:
+            print("SafeTensors not found in", model_name, e)
+            pass
+        print("num_params", model_name, num_params)
+
+    model_url = None
+    if still_on_hub:
+        model_url = f"https://huggingface.co/{model_name}"
+
+    model_license = "Unknown"
+    if model_config is not None:
+        info = model_info(repo_id=model_name)
+        # print(info.card_data)
+        model_license = info.card_data["license_name"]
+        model_license_link = info.card_data["license_link"]
+        if model_license_link:
+            model_license = f"[{model_license}]({model_license_link})"
+        if not model_license:
+            model_license = "Unknown"
+
+    return {
+        "model_architecture": architecture,
+        "model_type": "Open-Weights" if still_on_hub else "Proprietary",
+        "model_size": num_params if num_params else None,
+        "model_url": model_url,
+        "model_license": model_license,
+    }
+
+
+@dataclass
+class Field:
+    pretty_name: str
+    column_type: str
+    displayed_by_default: bool = True
+    never_hidden: bool = False
+    fully_hidden: bool = False
+    tags: list[str] = field(default_factory=list)
+
+
+MODEL_COLUMNS_DICT = {
+    "model_type_symbol": Field("T", "str", never_hidden=True),
+    "model_size_symbol": Field("S", "str", never_hidden=True),
+    "model_name": Field("Model Name", "markdown", never_hidden=True),
+    "model_type": Field("Type", "str", displayed_by_default=False),
+    "model_size": Field("#Params (B)", "number", displayed_by_default=False),
+    "model_architecture": Field("Architecture", "str", displayed_by_default=False),
+    "model_license": Field("License", "markdown", displayed_by_default=False),
+}
+
+U_MATH_COLUMNS_DICT = {
+    "rank": Field("Rank", "number", never_hidden=True),
+    **MODEL_COLUMNS_DICT,
+    "judge_model_name": Field("Judge Model Name", "markdown", displayed_by_default=False),
+    "u_math_acc": Field("U-MATH Acc", "number", never_hidden=True, tags=["u_math"]),
+    "u_math_text_acc": Field("U-MATH Text Acc", "number", tags=["u_math", "text"]),
+    "u_math_visual_acc": Field("U-MATH Visual Acc", "number", tags=["u_math", "visual"]),
+    "differential_calc_acc": Field("Diff Calc Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "differential_calc_text_acc": Field("Diff Calc Text Acc", "number", displayed_by_default=False, tags=["text"]),
+    "differential_calc_visual_acc": Field(
+        "Diff Calc Visual Acc", "number", displayed_by_default=False, tags=["visual"]
+    ),
+    "integral_calc_acc": Field("Integral Calc Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "integral_calc_text_acc": Field("Integral Calc Text Acc", "number", displayed_by_default=False, tags=["text"]),
+    "integral_calc_visual_acc": Field(
+        "Integral Calc Visual Acc", "number", displayed_by_default=False, tags=["visual"]
+    ),
+    "algebra_acc": Field("Algebra Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "algebra_text_acc": Field("Algebra Text Acc", "number", displayed_by_default=False, tags=["text"]),
+    "algebra_visual_acc": Field("Algebra Visual Acc", "number", displayed_by_default=False, tags=["visual"]),
+    "multivariable_calculus_acc": Field("Multivar Calc Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "multivariable_calculus_text_acc": Field(
+        "Multivar Calc Text Acc", "number", displayed_by_default=False, tags=["text"]
+    ),
+    "multivariable_calculus_visual_acc": Field(
+        "Multivar Calc Visual Acc", "number", displayed_by_default=False, tags=["visual"]
+    ),
+    "precalculus_review_acc": Field("Precalc Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "precalculus_review_text_acc": Field("Precalc Text Acc", "number", displayed_by_default=False, tags=["text"]),
+    "precalculus_review_visual_acc": Field(
+        "Precalc Visual Acc", "number", displayed_by_default=False, tags=["visual"]
+    ),
+    "sequences_series_acc": Field("Seq & Series Acc", "number", displayed_by_default=False, tags=["subjects"]),
+    "sequences_series_text_acc": Field("Seq & Series Text Acc", "number", displayed_by_default=False, tags=["text"]),
+    "sequences_series_visual_acc": Field(
+        "Seq & Series Visual Acc", "number", displayed_by_default=False, tags=["visual"]
+    ),
+}
+
+MU_MATH_COLUMNS_DICT = {
+    "rank": Field("Rank", "number", never_hidden=True),
+    **MODEL_COLUMNS_DICT,
+    "extract_model_name": Field("Extract Model Name", "markdown", displayed_by_default=False),
+    "mu_math_f1": Field("μ-MATH F1", "number", never_hidden=True, tags=["mu_math", "splits"]),
+    "mu_math_tpr": Field("μ-MATH TPR", "number", displayed_by_default=False, tags=["mu_math"]),
+    "mu_math_tnr": Field("μ-MATH TNR", "number", displayed_by_default=False, tags=["mu_math"]),
+    "mu_math_ppv": Field("μ-MATH PPV", "number", displayed_by_default=False, tags=["mu_math"]),
+    "mu_math_npv": Field("μ-MATH NPV", "number", displayed_by_default=False, tags=["mu_math"]),
+    "GPT-4o_f1": Field("GPT-4o Subset F1", "number", tags=["splits"]),
+    "GPT-4o_tpr": Field("GPT-4o Subset TPR", "number", displayed_by_default=False),
+    "GPT-4o_tnr": Field("GPT-4o Subset TNR", "number", displayed_by_default=False),
+    "GPT-4o_ppv": Field("GPT-4o Subset PPV", "number", displayed_by_default=False),
+    "GPT-4o_npv": Field("GPT-4o Subset NPV", "number", displayed_by_default=False),
+    "Gemini-1.5-Pro_f1": Field("Gemini-1.5-Pro Subset F1", "number", tags=["splits"]),
+    "Gemini-1.5-Pro_tpr": Field("Gemini-1.5-Pro Subset TPR", "number", displayed_by_default=False),
+    "Gemini-1.5-Pro_tnr": Field("Gemini-1.5-Pro Subset TNR", "number", displayed_by_default=False),
+    "Gemini-1.5-Pro_ppv": Field("Gemini-1.5-Pro Subset PPV", "number", displayed_by_default=False),
+    "Gemini-1.5-Pro_npv": Field("Gemini-1.5-Pro Subset NPV", "number", displayed_by_default=False),
+    "Llama-3.1-70B-Instruct_f1": Field("Llama-3.1-70B Subset F1", "number", tags=["splits"]),
+    "Llama-3.1-70B-Instruct_tpr": Field("Llama-3.1-70B Subset TPR", "number", displayed_by_default=False),
+    "Llama-3.1-70B-Instruct_tnr": Field("Llama-3.1-70B Subset TNR", "number", displayed_by_default=False),
+    "Llama-3.1-70B-Instruct_ppv": Field("Llama-3.1-70B Subset PPV", "number", displayed_by_default=False),
+    "Llama-3.1-70B-Instruct_npv": Field("Llama-3.1-70B Subset NPV", "number", displayed_by_default=False),
+    "Qwen2.5-72B-Instruct_f1": Field("Qwen2.5-72B Subset F1", "number", tags=["splits"]),
+    "Qwen2.5-72B-Instruct_tpr": Field("Qwen2.5-72B Subset TPR", "number", displayed_by_default=False),
+    "Qwen2.5-72B-Instruct_tnr": Field("Qwen2.5-72B Subset TNR", "number", displayed_by_default=False),
+    "Qwen2.5-72B-Instruct_ppv": Field("Qwen2.5-72B Subset PPV", "number", displayed_by_default=False),
+    "Qwen2.5-72B-Instruct_npv": Field("Qwen2.5-72B Subset NPV", "number", displayed_by_default=False),
+}
+
+
+def load_json_data(json_path: str, main_col: str | None = None) -> pd.DataFrame:
+    """Loads json data from a file"""
+    with open(json_path, "r") as f:
+        data = json.load(f)
+    df = pd.DataFrame.from_records(data)
+    if main_col:
+        df = df.sort_values(by=[main_col], ascending=False)
+
+    for col in df.columns:
+        if df.dtypes[col] == "float64":
+            df[col] = df[col].round(decimals=2)
     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]
+def get_u_math_leaderboard_df() -> pd.DataFrame:
+    """Creates a dataframe from json with U-MATH eval results"""
+    json_path = os.path.join("data", "u_math_eval_results.json")
+    df = load_json_data(json_path)
+
+    # flatten list [x, y, z] in columns as ["_acc", "_text_acc", "_visual_acc"] suffixes for columns
+    for col in [
+        "u_math",
+        "differential_calc",
+        "integral_calc",
+        "algebra",
+        "multivariable_calculus",
+        "precalculus_review",
+        "sequences_series",
+    ]:
+        df[col + "_acc"] = df[col].apply(lambda x: x[0])
+        df[col + "_text_acc"] = df[col].apply(lambda x: x[1])
+        df[col + "_visual_acc"] = df[col].apply(lambda x: x[2])
+        del df[col]
+
+    # Sort and add rank
+    df = df.sort_values(by=["u_math_acc"], ascending=False)
+    df["rank"] = range(1, len(df) + 1)
+
+    # populate with model info
+    model_to_meta_dict = {
+        model_name: get_hf_data_by_model_name(model_name) for model_name in df["model_name"].unique()
+    }
+    df["model_architecture"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_architecture"])
+    df["model_license"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_license"])
+    df["model_type"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_type"])
+    df["model_type_symbol"] = df["model_type"].apply(model_type_to_symbol)
+    df["model_size"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_size"])
+    df["model_size_symbol"] = df["model_size"].apply(model_size_to_symbol)
+    df["model_name"] = df["model_name"].apply(
+        lambda x: f"[{x}]({url})" if (url := model_to_meta_dict[x]["model_url"]) else x
+    )
+
+    # convert to pretty names and sort columns by order in dict
+    df = df[U_MATH_COLUMNS_DICT.keys()]
+    df = df.rename(columns={key: col.pretty_name for key, col in U_MATH_COLUMNS_DICT.items() if key in df.columns})
+
+    return df
+
+
+def get_mu_math_leaderboard_df() -> pd.DataFrame:
+    """Creates a dataframe from json with mu-MATH eval results"""
+    json_path = os.path.join("data", "mu_math_eval_results.json")
+    df = load_json_data(json_path)
+
+    # Calculate columns with prefixes f1, tpr, tnr, ppv, npv
+    for col in ["mu_math", "GPT-4o", "Gemini-1.5-Pro", "Llama-3.1-70B-Instruct", "Qwen2.5-72B-Instruct"]:
+        df[col + "_f1"] = df[col].apply(lambda x: x[0])
+        df[col + "_tpr"] = df[col].apply(lambda x: x[1])
+        df[col + "_tnr"] = df[col].apply(lambda x: x[2])
+        df[col + "_ppv"] = df[col].apply(lambda x: x[3])
+        df[col + "_npv"] = df[col].apply(lambda x: x[4])
+        del df[col]
+
+    # # flatten list [x, y, z] in columns as ["_f1", "_precision", "_recall"] suffixes for columns
+    # for col in ["mu_math", "GPT-4o", "Gemini-1.5-Pro", "Llama-3.1-70B-Instruct", "Qwen2.5-72B-Instruct"]:
+    #     df[col + "_f1"] = df[col].apply(lambda x: x[0])
+    #     df[col + "_precision"] = df[col].apply(lambda x: x[1])
+    #     df[col + "_recall"] = df[col].apply(lambda x: x[2])
+    #     del df[col]
+
+    NUM_MU_MATH_SAMPLES = 1084
+
+    # Sort and add rank
+    df = df.sort_values(by=["mu_math_f1"], ascending=False)
+    df["rank"] = range(1, len(df) + 1)
+
+    # populate with model info
+    model_to_meta_dict = {
+        model_name: get_hf_data_by_model_name(model_name) for model_name in df["model_name"].unique()
+    }
+    df["model_architecture"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_architecture"])
+    df["model_license"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_license"])
+    df["model_type"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_type"])
+    df["model_type_symbol"] = df["model_type"].apply(model_type_to_symbol)
+    df["model_size"] = df["model_name"].apply(lambda x: model_to_meta_dict[x]["model_size"])
+    df["model_size_symbol"] = df["model_size"].apply(model_size_to_symbol)
+    df["model_name"] = df["model_name"].apply(
+        lambda x: f"[{x}]({url})" if (url := model_to_meta_dict[x]["model_url"]) else x
+    )
+
+    # convert to pretty names and sort columns by order in dict
+    df = df[MU_MATH_COLUMNS_DICT.keys()]
+    df = df.rename(columns={key: col.pretty_name for key, col in MU_MATH_COLUMNS_DICT.items() if key in df.columns})
+
+    return df

src/submission/check_validity.py

@@ -1,99 +0,0 @@
-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

@@ -1,119 +0,0 @@
-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."
-    )