search-fix

.gitignore

@@ -1,19 +1,15 @@
+auto_evals/
 venv/
-.venv/
 __pycache__/
 .env
 .ipynb_checkpoints
 *ipynb
 .vscode/
-.DS_Store
-.ruff_cache/
-.python-version
-.profile_app.python
-*pstats
-poetry.lock
 
+gpt_4_evals/
+human_evals/
 eval-queue/
 eval-results/
-dynamic-info/
+auto_evals/
 
 src/assets/model_counts.html

README.md

@@ -4,21 +4,11 @@ emoji: 🏆
 colorFrom: green
 colorTo: indigo
 sdk: gradio
-sdk_version: 4.9.0
+sdk_version: 3.43.2
 app_file: app.py
 pinned: true
 license: apache-2.0
 duplicated_from: HuggingFaceH4/open_llm_leaderboard
-fullWidth: true
-startup_duration_timeout: 1h
-space_ci:
-  private: true
-  secrets:
-  - HF_TOKEN
-  - H4_TOKEN
-tags:
-- leaderboard
-short_description: Track, rank and evaluate open LLMs and chatbots
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,265 +1,291 @@
+import json
 import os
-import time
-import logging
+from datetime import datetime, timezone
+
 import gradio as gr
 import pandas as pd
 from apscheduler.schedulers.background import BackgroundScheduler
-from huggingface_hub import snapshot_download
-from gradio_space_ci import enable_space_ci
+from huggingface_hub import HfApi
 
-from src.display.about import (
+from src.assets.css_html_js import custom_css, get_window_url_params
+from src.assets.text_content import (
     CITATION_BUTTON_LABEL,
     CITATION_BUTTON_TEXT,
     EVALUATION_QUEUE_TEXT,
-    FAQ_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,
-    NUMERIC_INTERVALS,
-    TYPES,
+from src.display_models.get_model_metadata import DO_NOT_SUBMIT_MODELS, ModelType
+from src.display_models.modelcard_filter import check_model_card
+from src.display_models.utils import (
     AutoEvalColumn,
-    ModelType,
-    Precision,
-    WeightType,
+    EvalQueueColumn,
     fields,
+    styled_error,
+    styled_message,
+    styled_warning,
 )
-from src.envs import (
-    API,
-    DYNAMIC_INFO_FILE_PATH,
-    DYNAMIC_INFO_PATH,
-    DYNAMIC_INFO_REPO,
-    EVAL_REQUESTS_PATH,
-    EVAL_RESULTS_PATH,
-    H4_TOKEN,
-    IS_PUBLIC,
-    QUEUE_REPO,
-    REPO_ID,
-    RESULTS_REPO,
-)
-from src.populate import get_evaluation_queue_df, get_leaderboard_df
-from src.scripts.update_all_request_files import update_dynamic_files
-from src.submission.submit import add_new_eval
-from src.tools.collections import update_collections
-from src.tools.plots import create_metric_plot_obj, create_plot_df, create_scores_df
+from src.load_from_hub import get_evaluation_queue_df, get_leaderboard_df, is_model_on_hub, load_all_info_from_hub
+from src.rate_limiting import user_submission_permission
+
+pd.set_option("display.precision", 1)
+
+# clone / pull the lmeh eval data
+H4_TOKEN = os.environ.get("H4_TOKEN", None)
+
+QUEUE_REPO = "open-llm-leaderboard/requests"
+RESULTS_REPO = "open-llm-leaderboard/results"
 
+PRIVATE_QUEUE_REPO = "open-llm-leaderboard/private-requests"
+PRIVATE_RESULTS_REPO = "open-llm-leaderboard/private-results"
 
-# Configure logging
-logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
 
-# Start ephemeral Spaces on PRs (see config in README.md)
-enable_space_ci()
+EVAL_REQUESTS_PATH = "eval-queue"
+EVAL_RESULTS_PATH = "eval-results"
+
+EVAL_REQUESTS_PATH_PRIVATE = "eval-queue-private"
+EVAL_RESULTS_PATH_PRIVATE = "eval-results-private"
+
+api = HfApi(token=H4_TOKEN)
 
 
 def restart_space():
-    API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)
-
-
-def time_diff_wrapper(func):
-    def wrapper(*args, **kwargs):
-        start_time = time.time()
-        result = func(*args, **kwargs)
-        end_time = time.time()
-        diff = end_time - start_time
-        logging.info(f"Time taken for {func.__name__}: {diff} seconds")
-        return result
-    return wrapper
-
-
-@time_diff_wrapper
-def download_dataset(repo_id, local_dir, repo_type="dataset", max_attempts=3, backoff_factor=1.5):
-    """Download dataset with exponential backoff retries."""
-    attempt = 0
-    while attempt < max_attempts:
-        try:
-            logging.info(f"Downloading {repo_id} to {local_dir}")
-            snapshot_download(
-                repo_id=repo_id,
-                local_dir=local_dir,
-                repo_type=repo_type,
-                tqdm_class=None,
-                etag_timeout=30,
-                max_workers=8,
-            )
-            logging.info("Download successful")
-            return
-        except Exception as e:
-            wait_time = backoff_factor ** attempt
-            logging.error(f"Error downloading {repo_id}: {e}, retrying in {wait_time}s")
-            time.sleep(wait_time)
-            attempt += 1
-    raise Exception(f"Failed to download {repo_id} after {max_attempts} attempts")
-
-def init_space(full_init: bool = True):
-    """Initializes the application space, loading only necessary data."""
-    if full_init:
-        # These downloads only occur on full initialization
-        try:
-            download_dataset(QUEUE_REPO, EVAL_REQUESTS_PATH)
-            download_dataset(DYNAMIC_INFO_REPO, DYNAMIC_INFO_PATH)
-            download_dataset(RESULTS_REPO, EVAL_RESULTS_PATH)
-        except Exception:
-            restart_space()
-
-    # Always retrieve the leaderboard DataFrame
-    raw_data, original_df = get_leaderboard_df(
-        results_path=EVAL_RESULTS_PATH,
-        requests_path=EVAL_REQUESTS_PATH,
-        dynamic_path=DYNAMIC_INFO_FILE_PATH,
-        cols=COLS,
-        benchmark_cols=BENCHMARK_COLS,
-    )
+    api.restart_space(repo_id="HuggingFaceH4/open_llm_leaderboard", token=H4_TOKEN)
 
-    if full_init:
-        # Collection update only happens on full initialization
-        update_collections(original_df)
 
-    leaderboard_df = original_df.copy()
-    
-    # Evaluation queue DataFrame retrieval is independent of initialization detail level
-    eval_queue_dfs = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
+# Rate limit variables
+RATE_LIMIT_PERIOD = 7
+RATE_LIMIT_QUOTA = 5
 
-    return leaderboard_df, raw_data, original_df, eval_queue_dfs
+# Column selection
+COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
+TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
+COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
+TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
 
-# Convert the environment variable "LEADERBOARD_FULL_INIT" to a boolean value, defaulting to True if the variable is not set.
-# This controls whether a full initialization should be performed.
-do_full_init = os.getenv("LEADERBOARD_FULL_INIT", "True") == "True"
+if not IS_PUBLIC:
+    COLS.insert(2, AutoEvalColumn.precision.name)
+    TYPES.insert(2, AutoEvalColumn.precision.type)
 
-# Calls the init_space function with the `full_init` parameter determined by the `do_full_init` variable.
-# This initializes various DataFrames used throughout the application, with the level of initialization detail controlled by the `do_full_init` flag.
-leaderboard_df, raw_data, original_df, eval_queue_dfs = init_space(full_init=do_full_init)
-finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = eval_queue_dfs
+EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
+EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
 
+BENCHMARK_COLS = [
+    c.name
+    for c in [
+        AutoEvalColumn.arc,
+        AutoEvalColumn.hellaswag,
+        AutoEvalColumn.mmlu,
+        AutoEvalColumn.truthfulqa,
+    ]
+]
 
-# Data processing for plots now only on demand in the respective Gradio tab
-def load_and_create_plots():
-    plot_df = create_plot_df(create_scores_df(raw_data))
-    return plot_df
+## LOAD INFO FROM HUB
+eval_queue, requested_models, eval_results, users_to_submission_dates = load_all_info_from_hub(
+    QUEUE_REPO, RESULTS_REPO, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH
+)
+
+if not IS_PUBLIC:
+    (eval_queue_private, requested_models_private, eval_results_private, _) = load_all_info_from_hub(
+        PRIVATE_QUEUE_REPO,
+        PRIVATE_RESULTS_REPO,
+        EVAL_REQUESTS_PATH_PRIVATE,
+        EVAL_RESULTS_PATH_PRIVATE,
+    )
+else:
+    eval_queue_private, eval_results_private = None, None
+
+original_df = get_leaderboard_df(eval_results, eval_results_private, COLS, BENCHMARK_COLS)
+models = original_df["model_name_for_query"].tolist()  # needed for model backlinks in their to the leaderboard
+
+to_be_dumped = f"models = {repr(models)}\n"
+
+leaderboard_df = original_df.copy()
+(
+    finished_eval_queue_df,
+    running_eval_queue_df,
+    pending_eval_queue_df,
+) = get_evaluation_queue_df(eval_queue, eval_queue_private, EVAL_REQUESTS_PATH, EVAL_COLS)
+
+
+## INTERACTION FUNCTIONS
+def add_new_eval(
+    model: str,
+    base_model: str,
+    revision: str,
+    precision: str,
+    private: bool,
+    weight_type: str,
+    model_type: str,
+):
+    precision = precision.split(" ")[0]
+    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+    if model_type is None or model_type == "":
+        return styled_error("Please select a model type.")
+
+    # Is the user rate limited?
+    num_models_submitted_in_period = user_submission_permission(model, users_to_submission_dates, RATE_LIMIT_PERIOD)
+    if num_models_submitted_in_period > RATE_LIMIT_QUOTA:
+        error_msg = f"Organisation or user `{model.split('/')[0]}`"
+        error_msg += f"already has {num_models_submitted_in_period} model requests submitted to the leaderboard "
+        error_msg += f"in the last {RATE_LIMIT_PERIOD} days.\n"
+        error_msg += (
+            "Please wait a couple of days before resubmitting, so that everybody can enjoy using the leaderboard 🤗"
+        )
+        return styled_error(error_msg)
+
+    # Did the model authors forbid its submission to the leaderboard?
+    if model in DO_NOT_SUBMIT_MODELS or base_model in DO_NOT_SUBMIT_MODELS:
+        return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.")
+
+    # Does the model actually exist?
+    if revision == "":
+        revision = "main"
+
+    if weight_type in ["Delta", "Adapter"]:
+        base_model_on_hub, error = is_model_on_hub(base_model, revision)
+        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, revision)
+        if not model_on_hub:
+            return styled_error(f'Model "{model}" {error}')
+
+    # Were the model card and license filled?
+    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,
+        "private": private,
+        "precision": precision,
+        "weight_type": weight_type,
+        "status": "PENDING",
+        "submitted_time": current_time,
+        "model_type": model_type,
+    }
+
+    user_name = ""
+    model_path = model
+    if "/" in model:
+        user_name = model.split("/")[0]
+        model_path = model.split("/")[1]
+
+    print("Creating eval file")
+    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
+    os.makedirs(OUT_DIR, exist_ok=True)
+    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json"
+
+    # Check for duplicate submission
+    if f"{model}_{revision}_{precision}" in requested_models:
+        return styled_warning("This model has been already submitted.")
+
+    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."
+    )
+
+
+# Basics
+def change_tab(query_param: str):
+    query_param = query_param.replace("'", '"')
+    query_param = json.loads(query_param)
+
+    if isinstance(query_param, dict) and "tab" in query_param and query_param["tab"] == "evaluation":
+        return gr.Tabs.update(selected=1)
+    else:
+        return gr.Tabs.update(selected=0)
 
 
 # Searching and filtering
 def update_table(
     hidden_df: pd.DataFrame,
+    current_columns_df: pd.DataFrame,
     columns: list,
     type_query: list,
     precision_query: str,
     size_query: list,
-    hide_models: list,
+    show_deleted: bool,
     query: str,
 ):
-    filtered_df = filter_models(
-        df=hidden_df,
-        type_query=type_query,
-        size_query=size_query,
-        precision_query=precision_query,
-        hide_models=hide_models,
-    )
-    filtered_df = filter_queries(query, filtered_df)
+    filtered_df = filter_models(hidden_df, type_query, size_query, precision_query, show_deleted)
+    final_df = []
+    if query != "":
+        queries = query.split(";")
+        for _q in queries:
+            if _q != "":
+                temp_filtered_df = search_table(filtered_df, _q)
+                if len(temp_filtered_df) > 0:
+                    final_df.append(temp_filtered_df)
+        if len(final_df) > 0:
+            filtered_df = pd.concat(final_df).drop_duplicates()
     df = select_columns(filtered_df, columns)
     return df
 
 
-def load_query(request: gr.Request):  # triggered only once at startup => read query parameter if it exists
-    query = request.query_params.get("query") or ""
-    return (
-        query,
-        query,
-    )  # return one for the "search_bar", one for a hidden component that triggers a reload only if value has changed
-
+def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame:
+    return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))]
 
-def search_model(df: pd.DataFrame, query: str) -> pd.DataFrame:
-    return df[(df[AutoEvalColumn.fullname.name].str.contains(query, case=False, na=False))]
-
-def search_license(df: pd.DataFrame, query: str) -> pd.DataFrame:
-    return df[df[AutoEvalColumn.license.name].str.contains(query, case=False, na=False)]
 
 def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
-    always_here_cols = [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
-    dummy_col = [AutoEvalColumn.fullname.name]
-    filtered_df = df[always_here_cols + [c for c in COLS if c in df.columns and c in columns] + dummy_col]
-    return filtered_df
-
-def filter_queries(query: str, df: pd.DataFrame):
-    tmp_result_df = []
-
-    # Empty query return the same df
-    if query == "":
-        return df
-
-    # all_queries = [q.strip() for q in query.split(";")]
-    # license_queries = []
-    all_queries = [q.strip() for q in query.split(";") if q.strip() != ""]
-    model_queries = [q for q in all_queries if not q.startswith("licence")]
-    license_queries_raw = [q for q in all_queries if q.startswith("license")]
-    license_queries = [
-        q.replace("license:", "").strip() for q in license_queries_raw if q.replace("license:", "").strip() != ""
+    always_here_cols = [
+        AutoEvalColumn.model_type_symbol.name,
+        AutoEvalColumn.model.name,
     ]
+    # We use COLS to maintain sorting
+    filtered_df = df[
+        always_here_cols + [c for c in COLS if c in df.columns and c in columns] + [AutoEvalColumn.dummy.name]
+    ]
+    return filtered_df
 
-    # Handling model name search
-    for query in model_queries:
-        tmp_df = search_model(df, query)
-        if len(tmp_df) > 0:
-            tmp_result_df.append(tmp_df)
-
-    if not tmp_result_df and not license_queries:
-        # Nothing is found, no license_queries -> return empty df
-        return pd.DataFrame(columns=df.columns)
-
-    if tmp_result_df:
-        df = pd.concat(tmp_result_df)
-        df = df.drop_duplicates(
-            subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name]
-        )
-
-    if not license_queries:
-        return df
-
-    # Handling license search
-    tmp_result_df = []
-    for query in license_queries:
-        tmp_df = search_license(df, query)
-        if len(tmp_df) > 0:
-            tmp_result_df.append(tmp_df)
 
-    if not tmp_result_df:
-        # Nothing is found, return empty df
-        return pd.DataFrame(columns=df.columns)
-
-    df = pd.concat(tmp_result_df)
-    df = df.drop_duplicates(
-        subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name]
-    )
-
-    return df
+NUMERIC_INTERVALS = {
+    "Unknown": pd.Interval(-1, 0, closed="right"),
+    "< 1.5B": pd.Interval(0, 1.5, closed="right"),
+    "~3B": pd.Interval(1.5, 5, closed="right"),
+    "~7B": pd.Interval(6, 11, closed="right"),
+    "~13B": pd.Interval(12, 15, closed="right"),
+    "~35B": pd.Interval(16, 55, closed="right"),
+    "60B+": pd.Interval(55, 10000, closed="right"),
+}
 
 
 def filter_models(
-    df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, hide_models: list
+    df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool
 ) -> pd.DataFrame:
     # Show all models
-    if "Private or deleted" in hide_models:
-        filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
-    else:
+    if show_deleted:
         filtered_df = df
-
-    if "Contains a merge/moerge" in hide_models:
-        filtered_df = filtered_df[filtered_df[AutoEvalColumn.merged.name] == False]
-
-    if "MoE" in hide_models:
-        filtered_df = filtered_df[filtered_df[AutoEvalColumn.moe.name] == False]
-
-    if "Flagged" in hide_models:
-        filtered_df = filtered_df[filtered_df[AutoEvalColumn.flagged.name] == False]
+    else:  # Show only still on the hub models
+        filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
 
     type_emoji = [t[0] for t in type_query]
-    filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)]
-    filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])]
+    filtered_df = filtered_df[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)]
+    filtered_df = filtered_df[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])]
 
     numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query]))
     params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce")
@@ -269,14 +295,6 @@ def filter_models(
     return filtered_df
 
 
-leaderboard_df = filter_models(
-    df=leaderboard_df,
-    type_query=[t.to_str(" : ") for t in ModelType],
-    size_query=list(NUMERIC_INTERVALS.keys()),
-    precision_query=[i.value.name for i in Precision],
-    hide_models=["Private or deleted", "Contains a merge/moerge", "Flagged"],  # Deleted, merges, flagged, MoEs
-)
-
 demo = gr.Blocks(css=custom_css)
 with demo:
     gr.HTML(TITLE)
@@ -288,65 +306,92 @@ with demo:
                 with gr.Column():
                     with gr.Row():
                         search_bar = gr.Textbox(
-                            placeholder="🔍 Search models or licenses (e.g., 'model_name; license: MIT') and press ENTER...",
+                            placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...",
                             show_label=False,
                             elem_id="search-bar",
                         )
                     with gr.Row():
                         shown_columns = gr.CheckboxGroup(
                             choices=[
-                                c.name
-                                for c in fields(AutoEvalColumn)
-                                if not c.hidden and not c.never_hidden and not c.dummy
+                                c
+                                for c in COLS
+                                if c
+                                not in [
+                                    AutoEvalColumn.dummy.name,
+                                    AutoEvalColumn.model.name,
+                                    AutoEvalColumn.model_type_symbol.name,
+                                    AutoEvalColumn.still_on_hub.name,
+                                ]
                             ],
                             value=[
-                                c.name
-                                for c in fields(AutoEvalColumn)
-                                if c.displayed_by_default and not c.hidden and not c.never_hidden
+                                c
+                                for c in COLS_LITE
+                                if c
+                                not in [
+                                    AutoEvalColumn.dummy.name,
+                                    AutoEvalColumn.model.name,
+                                    AutoEvalColumn.model_type_symbol.name,
+                                    AutoEvalColumn.still_on_hub.name,
+                                ]
                             ],
                             label="Select columns to show",
                             elem_id="column-select",
                             interactive=True,
                         )
                     with gr.Row():
-                        hide_models = gr.CheckboxGroup(
-                            label="Hide models",
-                            choices=["Private or deleted", "Contains a merge/moerge", "Flagged", "MoE"],
-                            value=["Private or deleted", "Contains a merge/moerge", "Flagged"],
-                            interactive=True,
+                        deleted_models_visibility = gr.Checkbox(
+                            value=True, label="Show gated/private/deleted models", interactive=True
                         )
                 with gr.Column(min_width=320):
-                    # with gr.Box(elem_id="box-filter"):
-                    filter_columns_type = gr.CheckboxGroup(
-                        label="Model types",
-                        choices=[t.to_str() for t in ModelType],
-                        value=[t.to_str() for t in ModelType],
-                        interactive=True,
-                        elem_id="filter-columns-type",
-                    )
-                    filter_columns_precision = gr.CheckboxGroup(
-                        label="Precision",
-                        choices=[i.value.name for i in Precision],
-                        value=[i.value.name for i in Precision],
-                        interactive=True,
-                        elem_id="filter-columns-precision",
-                    )
-                    filter_columns_size = gr.CheckboxGroup(
-                        label="Model sizes (in billions of parameters)",
-                        choices=list(NUMERIC_INTERVALS.keys()),
-                        value=list(NUMERIC_INTERVALS.keys()),
-                        interactive=True,
-                        elem_id="filter-columns-size",
-                    )
+                    with gr.Box(elem_id="box-filter"):
+                        filter_columns_type = gr.CheckboxGroup(
+                            label="Model types",
+                            choices=[
+                                ModelType.PT.to_str(),
+                                ModelType.FT.to_str(),
+                                ModelType.IFT.to_str(),
+                                ModelType.RL.to_str(),
+                                ModelType.Unknown.to_str(),
+                            ],
+                            value=[
+                                ModelType.PT.to_str(),
+                                ModelType.FT.to_str(),
+                                ModelType.IFT.to_str(),
+                                ModelType.RL.to_str(),
+                                ModelType.Unknown.to_str(),
+                            ],
+                            interactive=True,
+                            elem_id="filter-columns-type",
+                        )
+                        filter_columns_precision = gr.CheckboxGroup(
+                            label="Precision",
+                            choices=["torch.float16", "torch.bfloat16", "torch.float32", "8bit", "4bit", "GPTQ"],
+                            value=["torch.float16", "torch.bfloat16", "torch.float32", "8bit", "4bit", "GPTQ"],
+                            interactive=True,
+                            elem_id="filter-columns-precision",
+                        )
+                        filter_columns_size = gr.CheckboxGroup(
+                            label="Model sizes",
+                            choices=list(NUMERIC_INTERVALS.keys()),
+                            value=list(NUMERIC_INTERVALS.keys()),
+                            interactive=True,
+                            elem_id="filter-columns-size",
+                        )
 
             leaderboard_table = gr.components.Dataframe(
                 value=leaderboard_df[
-                    [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
+                    [AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name]
                     + shown_columns.value
-                    + [AutoEvalColumn.fullname.name]
+                    + [AutoEvalColumn.dummy.name]
                 ],
-                headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value,
+                headers=[
+                    AutoEvalColumn.model_type_symbol.name,
+                    AutoEvalColumn.model.name,
+                ]
+                + shown_columns.value
+                + [AutoEvalColumn.dummy.name],
                 datatype=TYPES,
+                max_rows=None,
                 elem_id="leaderboard-table",
                 interactive=False,
                 visible=True,
@@ -354,121 +399,175 @@ with demo:
 
             # Dummy leaderboard for handling the case when the user uses backspace key
             hidden_leaderboard_table_for_search = gr.components.Dataframe(
-                value=original_df[COLS],
+                value=original_df,
                 headers=COLS,
                 datatype=TYPES,
+                max_rows=None,
                 visible=False,
             )
             search_bar.submit(
                 update_table,
                 [
                     hidden_leaderboard_table_for_search,
+                    leaderboard_table,
                     shown_columns,
                     filter_columns_type,
                     filter_columns_precision,
                     filter_columns_size,
-                    hide_models,
+                    deleted_models_visibility,
                     search_bar,
                 ],
                 leaderboard_table,
             )
-
-            # Define a hidden component that will trigger a reload only if a query parameter has been set
-            hidden_search_bar = gr.Textbox(value="", visible=False)
-            hidden_search_bar.change(
+            shown_columns.change(
                 update_table,
                 [
                     hidden_leaderboard_table_for_search,
+                    leaderboard_table,
                     shown_columns,
                     filter_columns_type,
                     filter_columns_precision,
                     filter_columns_size,
-                    hide_models,
+                    deleted_models_visibility,
                     search_bar,
                 ],
                 leaderboard_table,
+                queue=True,
             )
-            # Check query parameter once at startup and update search bar + hidden component
-            demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar])
-
-            for selector in [
-                shown_columns,
-                filter_columns_type,
-                filter_columns_precision,
-                filter_columns_size,
-                hide_models,
-            ]:
-                selector.change(
-                    update_table,
-                    [
-                        hidden_leaderboard_table_for_search,
-                        shown_columns,
-                        filter_columns_type,
-                        filter_columns_precision,
-                        filter_columns_size,
-                        hide_models,
-                        search_bar,
-                    ],
+            filter_columns_type.change(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
                     leaderboard_table,
-                    queue=True,
-                )
-
-        with gr.TabItem("📈 Metrics through time", elem_id="llm-benchmark-tab-table", id=2):
-            with gr.Row():
-                with gr.Column():
-                    plot_df = load_and_create_plots()
-                    chart = create_metric_plot_obj(
-                        plot_df,
-                        [AutoEvalColumn.average.name],
-                        title="Average of Top Scores and Human Baseline Over Time (from last update)",
-                    )
-                    gr.Plot(value=chart, min_width=500)
-                with gr.Column():
-                    plot_df = load_and_create_plots()
-                    chart = create_metric_plot_obj(
-                        plot_df,
-                        BENCHMARK_COLS,
-                        title="Top Scores and Human Baseline Over Time (from last update)",
-                    )
-                    gr.Plot(value=chart, min_width=500)
-
-        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=3):
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    deleted_models_visibility,
+                    search_bar,
+                ],
+                leaderboard_table,
+                queue=True,
+            )
+            filter_columns_precision.change(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
+                    leaderboard_table,
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    deleted_models_visibility,
+                    search_bar,
+                ],
+                leaderboard_table,
+                queue=True,
+            )
+            filter_columns_size.change(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
+                    leaderboard_table,
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    deleted_models_visibility,
+                    search_bar,
+                ],
+                leaderboard_table,
+                queue=True,
+            )
+            deleted_models_visibility.change(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
+                    leaderboard_table,
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    deleted_models_visibility,
+                    search_bar,
+                ],
+                leaderboard_table,
+                queue=True,
+            )
+        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
             gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
 
-        with gr.TabItem("❗FAQ", elem_id="llm-benchmark-tab-table", id=4):
-            gr.Markdown(FAQ_TEXT, elem_classes="markdown-text")
-
-        with gr.TabItem("🚀 Submit ", elem_id="llm-benchmark-tab-table", id=5):
+        with gr.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,
+                                max_rows=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,
+                                max_rows=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,
+                                max_rows=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")
+                    revision_name_textbox = gr.Textbox(label="revision", placeholder="main")
                     private = gr.Checkbox(False, label="Private", visible=not IS_PUBLIC)
                     model_type = gr.Dropdown(
-                        choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
+                        choices=[
+                            ModelType.PT.to_str(" : "),
+                            ModelType.FT.to_str(" : "),
+                            ModelType.IFT.to_str(" : "),
+                            ModelType.RL.to_str(" : "),
+                        ],
                         label="Model type",
                         multiselect=False,
-                        value=ModelType.FT.to_str(" : "),
+                        value=None,
                         interactive=True,
                     )
 
                 with gr.Column():
                     precision = gr.Dropdown(
-                        choices=[i.value.name for i in Precision if i != Precision.Unknown],
+                        choices=["float16", "bfloat16", "8bit (LLM.int8)", "4bit (QLoRA / FP4)", "GPTQ"],
                         label="Precision",
                         multiselect=False,
                         value="float16",
                         interactive=True,
                     )
                     weight_type = gr.Dropdown(
-                        choices=[i.value.name for i in WeightType],
+                        choices=["Original", "Delta", "Adapter"],
                         label="Weights type",
                         multiselect=False,
                         value="Original",
@@ -476,42 +575,6 @@ with demo:
                     )
                     base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
 
-            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,
-                        )
-
             submit_button = gr.Button("Submit Eval")
             submission_result = gr.Markdown()
             submit_button.click(
@@ -535,12 +598,17 @@ with demo:
                 label=CITATION_BUTTON_LABEL,
                 lines=20,
                 elem_id="citation-button",
-                show_copy_button=True,
-            )
+            ).style(show_copy_button=True)
+
+    dummy = gr.Textbox(visible=False)
+    demo.load(
+        change_tab,
+        dummy,
+        tabs,
+        _js=get_window_url_params,
+    )
 
 scheduler = BackgroundScheduler()
-scheduler.add_job(restart_space, "interval", hours=3)  # restarted every 3h
-scheduler.add_job(update_dynamic_files, "interval", hours=2)  # launched every 2 hour
+scheduler.add_job(restart_space, "interval", seconds=1800)
 scheduler.start()
-
-demo.queue(default_concurrency_limit=40).launch()
+demo.queue(concurrency_count=40).launch()

model_info_cache.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fcaa2a3e1ac6a5559471547af5de4e3ccd49673ad5525890726e65cd90cfe62
+size 3620752

model_size_cache.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75d1f64589459eb64e3a50987bf05ed3656248102d1fe2f6c98a008020945840
+size 74321

pyproject.toml

@@ -1,15 +1,9 @@
 [tool.ruff]
-line-length = 120
-target-version = "py312"
-include = ["*.py", "*.pyi", "**/pyproject.toml", "*.ipynb"]
-ignore=["I","EM","FBT","TRY003","S101","D101","D102","D103","D104","D105","G004","D107","FA102"]
-fixable=["ALL"]
-select=["ALL"]
-
-[tool.ruff.lint]
+# Enable pycodestyle (`E`) and Pyflakes (`F`) codes by default.
 select = ["E", "F"]
-fixable = ["ALL"]
 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]
 profile = "black"
@@ -17,37 +11,3 @@ line_length = 119
 
 [tool.black]
 line-length = 119
-
-[tool.poetry]
-package-mode = false
-name = "open-llm-leaderboard"
-version = "0.1.0"
-description = ""
-authors = []
-readme = "README.md"
-
-[tool.poetry.dependencies]
-python = "3.12.1"
-apscheduler = "3.10.1"
-black = "23.11.0"
-click = "8.1.3"
-datasets = "2.14.5"
-huggingface-hub = ">=0.18.0"
-matplotlib = "3.8.4"
-numpy = "1.26.0"
-pandas = "2.2.2"
-plotly = "5.14.1"
-python-dateutil = "2.8.2"
-requests = "2.28.2"
-sentencepiece = "^0.2.0"
-tqdm = "4.65.0"
-transformers = "4.40.0"
-tokenizers = ">=0.15.0"
-gradio-space-ci = {git = "https://huggingface.co/spaces/Wauplin/gradio-space-ci", rev = "0.2.3"}
-gradio = "4.9.0"
-isort = "^5.13.2"
-ruff = "^0.3.5"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"

requirements.txt

@@ -1,16 +1,71 @@
+accelerate==0.23.0
+aiofiles==23.1.0
+aiohttp==3.8.4
+aiosignal==1.3.1
+altair==4.2.2
+anyio==3.6.2
 APScheduler==3.10.1
-black==23.11.0
+async-timeout==4.0.2
+attrs==23.1.0
+certifi==2022.12.7
+charset-normalizer==3.1.0
 click==8.1.3
-datasets==2.14.5
-huggingface-hub>=0.18.0
-matplotlib==3.8.4
-numpy==1.26.0
-pandas==2.2.2
+contourpy==1.0.7
+cycler==0.11.0
+datasets==2.12.0
+entrypoints==0.4
+fastapi==0.95.1
+ffmpy==0.3.0
+filelock==3.11.0
+fonttools==4.39.3
+frozenlist==1.3.3
+fsspec==2023.4.0
+gradio==3.43.2
+gradio-client==0.5.0
+h11==0.14.0
+httpcore==0.17.0
+httpx==0.24.0
+huggingface-hub==0.16.4
+idna==3.4
+Jinja2==3.1.2
+jsonschema==4.17.3
+kiwisolver==1.4.4
+linkify-it-py==2.0.0
+markdown-it-py==2.2.0
+MarkupSafe==2.1.2
+matplotlib==3.7.1
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+multidict==6.0.4
+numpy==1.24.2
+orjson==3.8.10
+packaging==23.1
+pandas==2.0.0
+Pillow==9.5.0
 plotly==5.14.1
+pyarrow==11.0.0
+pydantic==1.10.7
+pydub==0.25.1
+pyparsing==3.0.9
+pyrsistent==0.19.3
 python-dateutil==2.8.2
+python-multipart==0.0.6
+pytz==2023.3
+pytz-deprecation-shim==0.1.0.post0
+PyYAML==6.0
 requests==2.28.2
-sentencepiece
+semantic-version==2.10.0
+six==1.16.0
+sniffio==1.3.0
+starlette==0.26.1
+toolz==0.12.0
 tqdm==4.65.0
-transformers==4.40.0
-tokenizers>=0.15.0
-gradio-space-ci @ git+https://huggingface.co/spaces/Wauplin/gradio-space-ci@0.2.3 # CI !!!
+transformers@git+https://github.com/huggingface/transformers
+typing_extensions==4.5.0
+tzdata==2023.3
+tzlocal==4.3
+uc-micro-py==1.0.1
+urllib3==1.26.15
+uvicorn==0.21.1
+websockets==11.0.1
+yarl==1.8.2

src/{display → assets}/css_html_js.py

@@ -1,25 +1,5 @@
 custom_css = """
 
-/* Hides the final AutoEvalColumn */
-#llm-benchmark-tab-table table td:last-child,
-#llm-benchmark-tab-table table th:last-child {
-    display: none;
-}
-
-/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
-table td:first-child,
-table th:first-child {
-    max-width: 400px;
-    overflow: auto;
-    white-space: nowrap;
-}
-
-/* Full width space */
-.gradio-container {
-  max-width: 95%!important;
-}
-
-/* Text style and margins */
 .markdown-text {
     font-size: 16px !important;
 }
@@ -41,20 +21,53 @@ table th:first-child {
     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;
 }
 
+/* Hides the final AutoEvalColumn */
+#llm-benchmark-tab-table table td:last-child,
+#llm-benchmark-tab-table table th:last-child {
+    display: none;
+}
+
+/* Limit the width of the first AutoEvalColumn so that names don't expand too much */
+table td:first-child,
+table th:first-child {
+    max-width: 400px;
+    overflow: auto;
+    white-space: nowrap;
+}
+
 .tab-buttons button {
     font-size: 20px;
 }
 
-/* Filters style */
+#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;
@@ -95,4 +108,4 @@ get_window_url_params = """
         url_params = Object.fromEntries(params);
         return url_params;
     }
-    """
+    """

src/assets/hardcoded_evals.py

@@ -0,0 +1,40 @@
+from src.display_models.utils import AutoEvalColumn, model_hyperlink
+
+gpt4_values = {
+    AutoEvalColumn.model.name: model_hyperlink("https://arxiv.org/abs/2303.08774", "gpt4"),
+    AutoEvalColumn.revision.name: "tech report",
+    AutoEvalColumn.precision.name: None,
+    AutoEvalColumn.average.name: 84.3,
+    AutoEvalColumn.arc.name: 96.3,
+    AutoEvalColumn.hellaswag.name: 95.3,
+    AutoEvalColumn.mmlu.name: 86.4,
+    AutoEvalColumn.truthfulqa.name: 59.0,
+    AutoEvalColumn.dummy.name: "GPT-4",
+    AutoEvalColumn.model_type.name: "",
+}
+
+gpt35_values = {
+    AutoEvalColumn.model.name: model_hyperlink("https://arxiv.org/abs/2303.08774", "gpt3.5"),
+    AutoEvalColumn.revision.name: "tech report",
+    AutoEvalColumn.precision.name: None,
+    AutoEvalColumn.average.name: 71.9,
+    AutoEvalColumn.arc.name: 85.2,
+    AutoEvalColumn.hellaswag.name: 85.5,
+    AutoEvalColumn.mmlu.name: 70.0,
+    AutoEvalColumn.truthfulqa.name: 47.0,
+    AutoEvalColumn.dummy.name: "GPT-3.5",
+    AutoEvalColumn.model_type.name: "",
+}
+
+baseline = {
+    AutoEvalColumn.model.name: "<p>Baseline</p>",
+    AutoEvalColumn.revision.name: "N/A",
+    AutoEvalColumn.precision.name: None,
+    AutoEvalColumn.average.name: 25.0,
+    AutoEvalColumn.arc.name: 25.0,
+    AutoEvalColumn.hellaswag.name: 25.0,
+    AutoEvalColumn.mmlu.name: 25.0,
+    AutoEvalColumn.truthfulqa.name: 25.0,
+    AutoEvalColumn.dummy.name: "baseline",
+    AutoEvalColumn.model_type.name: "",
+}

src/{display/about.py → assets/text_content.py}

@@ -1,62 +1,52 @@
-from src.display.utils import ModelType
+from src.display_models.model_metadata_type import ModelType
 
-TITLE = """<h1 style="text-align:left;float:left; id="space-title">🤗 Open LLM Leaderboard</h1> <h3 style="text-align:left;float:left;> Track, rank and evaluate open LLMs and chatbots </h3>"""
+TITLE = """<h1 align="center" id="space-title">🤗 Open LLM Leaderboard</h1>"""
 
 INTRODUCTION_TEXT = """
-"""
+📐 The 🤗 Open LLM Leaderboard aims to track, rank and evaluate open LLMs and chatbots.
 
-icons = f"""
-- {ModelType.PT.to_str(" : ")} model: new, base models, trained on a given text corpora using masked modelling
-- {ModelType.CPT.to_str(" : ")} model: new, base models, continuously trained on further corpus (which may include IFT/chat data) using masked modelling
-- {ModelType.FT.to_str(" : ")} model: pretrained models finetuned on more data
-- {ModelType.chat.to_str(" : ")} model: chat like fine-tunes, either using IFT (datasets of task instruction), RLHF or DPO (changing the model loss a bit with an added policy), etc
-- {ModelType.merges.to_str(" : ")} model: merges or MoErges, models which have been merged or fused without additional fine-tuning. 
+🤗 Submit a model for automated evaluation on the 🤗 GPU cluster on the "Submit" page!
+The leaderboard's backend runs the great [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) - read more details in the "About" page!
 """
-LLM_BENCHMARKS_TEXT = """
-## ABOUT
+
+LLM_BENCHMARKS_TEXT = f"""
+# Context
 With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art.
 
-🤗 Submit a model for automated evaluation on the 🤗 GPU cluster on the "Submit" page!
-The leaderboard's backend runs the great [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) - read more details below!
+## Icons
+{ModelType.PT.to_str(" : ")} model
+{ModelType.FT.to_str(" : ")} model
+{ModelType.IFT.to_str(" : ")} model
+{ModelType.RL.to_str(" : ")} model
+If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
 
-### Tasks 
-📈 We evaluate models on 6 key benchmarks using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks.
+🏴‍☠️ indicates that this model has been flagged by the community, and should probably be ignored! Clicking the icon will redirect you to the discussion about the model.
+(For ex, the model was trained on the evaluation data, and is therefore cheating on the leaderboard.)
+
+## How it works
+
+📈 We evaluate models on 4 key benchmarks using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks.
 
 - <a href="https://arxiv.org/abs/1803.05457" target="_blank">  AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions.
 - <a href="https://arxiv.org/abs/1905.07830" target="_blank">  HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
 - <a href="https://arxiv.org/abs/2009.03300" target="_blank">  MMLU </a>  (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
-- <a href="https://arxiv.org/abs/2109.07958" target="_blank">  TruthfulQA </a> (0-shot) - a test to measure a model's propensity to reproduce falsehoods commonly found online. Note: TruthfulQA is technically a 6-shot task in the Harness because each example is prepended with 6 Q/A pairs, even in the 0-shot setting.
-- <a href="https://arxiv.org/abs/1907.10641" target="_blank">  Winogrande </a> (5-shot) - an adversarial and difficult Winograd benchmark at scale, for commonsense reasoning.
-- <a href="https://arxiv.org/abs/2110.14168" target="_blank">  GSM8k </a> (5-shot) - diverse grade school math word problems to measure a model's ability to solve multi-step mathematical reasoning problems.
+- <a href="https://arxiv.org/abs/2109.07958" target="_blank">  TruthfulQA </a> (0-shot) - a test to measure a model’s propensity to reproduce falsehoods commonly found online. Note: TruthfulQA in the Harness is actually a minima a 6-shots task, as it is prepended by 6 examples systematically, even when launched using 0 for the number of few-shot examples.
 
 For all these evaluations, a higher score is a better score.
 We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings.
 
-### Results
+## Details and logs
 You can find:
 - detailed numerical results in the `results` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/results
-- details on the input/outputs for the models in the `details` of each model, which you can access by clicking the 📄 emoji after the model name
+- details on the input/outputs for the models in the `details` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/details
 - community queries and running status in the `requests` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/requests
 
-If a model's name contains "Flagged", this indicates it has been flagged by the community, and should probably be ignored! Clicking the link will redirect you to the discussion about the model.
-
----------------------------
+## Reproducibility
+To reproduce our results, here is the commands you can run, using [this version](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463) of the Eleuther AI Harness:
+`python main.py --model=hf-causal --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>"`
+` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=2 --output_path=<output_path>`
 
-## REPRODUCIBILITY
-To reproduce our results, here are the commands you can run, using [this version](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463) of the Eleuther AI Harness:
-`python main.py --model=hf-causal-experimental --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>"`
-` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=1 --output_path=<output_path>`
-
-```
-python main.py --model=hf-causal-experimental \
-    --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>" \
-    --tasks=<task_list> \
-    --num_fewshot=<n_few_shot> \
-    --batch_size=1 \
-    --output_path=<output_path>
-```
-
-**Note:** We evaluate all models on a single node of 8 H100s, so the global batch size is 8 for each evaluation. If you don't use parallelism, adapt your batch size to fit.
+The total batch size we get for models which fit on one A100 node is 16 (8 GPUs * 2). If you don't use parallelism, adapt your batch size to fit.
 *You can expect results to vary slightly for different batch sizes because of padding.*
 
 The tasks and few shots parameters are:
@@ -64,122 +54,23 @@ The tasks and few shots parameters are:
 - HellaSwag: 10-shot, *hellaswag* (`acc_norm`)
 - TruthfulQA: 0-shot, *truthfulqa-mc* (`mc2`)
 - MMLU: 5-shot, *hendrycksTest-abstract_algebra,hendrycksTest-anatomy,hendrycksTest-astronomy,hendrycksTest-business_ethics,hendrycksTest-clinical_knowledge,hendrycksTest-college_biology,hendrycksTest-college_chemistry,hendrycksTest-college_computer_science,hendrycksTest-college_mathematics,hendrycksTest-college_medicine,hendrycksTest-college_physics,hendrycksTest-computer_security,hendrycksTest-conceptual_physics,hendrycksTest-econometrics,hendrycksTest-electrical_engineering,hendrycksTest-elementary_mathematics,hendrycksTest-formal_logic,hendrycksTest-global_facts,hendrycksTest-high_school_biology,hendrycksTest-high_school_chemistry,hendrycksTest-high_school_computer_science,hendrycksTest-high_school_european_history,hendrycksTest-high_school_geography,hendrycksTest-high_school_government_and_politics,hendrycksTest-high_school_macroeconomics,hendrycksTest-high_school_mathematics,hendrycksTest-high_school_microeconomics,hendrycksTest-high_school_physics,hendrycksTest-high_school_psychology,hendrycksTest-high_school_statistics,hendrycksTest-high_school_us_history,hendrycksTest-high_school_world_history,hendrycksTest-human_aging,hendrycksTest-human_sexuality,hendrycksTest-international_law,hendrycksTest-jurisprudence,hendrycksTest-logical_fallacies,hendrycksTest-machine_learning,hendrycksTest-management,hendrycksTest-marketing,hendrycksTest-medical_genetics,hendrycksTest-miscellaneous,hendrycksTest-moral_disputes,hendrycksTest-moral_scenarios,hendrycksTest-nutrition,hendrycksTest-philosophy,hendrycksTest-prehistory,hendrycksTest-professional_accounting,hendrycksTest-professional_law,hendrycksTest-professional_medicine,hendrycksTest-professional_psychology,hendrycksTest-public_relations,hendrycksTest-security_studies,hendrycksTest-sociology,hendrycksTest-us_foreign_policy,hendrycksTest-virology,hendrycksTest-world_religions* (average of all the results `acc`)
-- Winogrande: 5-shot, *winogrande* (`acc`)
-- GSM8k: 5-shot, *gsm8k* (`acc`)
-
-Side note on the baseline scores: 
-- for log-likelihood evaluation, we select the random baseline
-- for GSM8K, we select the score obtained in the paper after finetuning a 6B model on the full GSM8K training set for 50 epochs
-
----------------------------
-
-## RESOURCES
 
-### Quantization
+## Quantization
 To get more information about quantization, see:
 - 8 bits: [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration), [paper](https://arxiv.org/abs/2208.07339)
 - 4 bits: [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes), [paper](https://arxiv.org/abs/2305.14314)
 
-### Useful links
-- [Community resources](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/174)
-- [Collection of best models](https://huggingface.co/collections/open-llm-leaderboard/llm-leaderboard-best-models-652d6c7965a4619fb5c27a03)
-
-### Other cool leaderboards:
-- [LLM safety](https://huggingface.co/spaces/AI-Secure/llm-trustworthy-leaderboard)
-- [LLM performance](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)
-
-
+## More resources
+If you still have questions, you can check our FAQ [here](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/179)!
+We also gather cool resources from the community, other teams, and other labs [here](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/174)!
 """
 
-FAQ_TEXT = """
-
-## SUBMISSIONS
-My model requires `trust_remote_code=True`, can I submit it?
-- *We only support models that have been integrated into a stable version of the `transformers` library for automatic submission, as we don't want to run possibly unsafe code on our cluster.*
-
-What about models of type X? 
-- *We only support models that have been integrated into a stable version of the `transformers` library for automatic submission.*
-
-How can I follow when my model is launched?
-- *You can look for its request file [here](https://huggingface.co/datasets/open-llm-leaderboard/requests) and follow the status evolution, or directly in the queues above the submit form.*
-
-My model disappeared from all the queues, what happened?
-- *A model disappearing from all the queues usually means that there has been a failure. You can check if that is the case by looking for your model [here](https://huggingface.co/datasets/open-llm-leaderboard/requests).*
-
-What causes an evaluation failure?
-- *Most of the failures we get come from problems in the submissions (corrupted files, config problems, wrong parameters selected for eval ...), so we'll be grateful if you first make sure you have followed the steps in `About`. However, from time to time, we have failures on our side (hardware/node failures, problems with an update of our backend, connectivity problems ending up in the results not being saved, ...).*
-
-How can I report an evaluation failure?
-- *As we store the logs for all models, feel free to create an issue, **where you link to the requests file of your model** (look for it [here](https://huggingface.co/datasets/open-llm-leaderboard/requests/tree/main)), so we can investigate! If the model failed due to a problem on our side, we'll relaunch it right away!* 
-*Note: Please do not re-upload your model under a different name, it will not help*
-
----------------------------
-
-## RESULTS
-What kind of information can I find?
-- *Let's imagine you are interested in the Yi-34B results. You have access to 3 different information categories:*
-      - *The [request file](https://huggingface.co/datasets/open-llm-leaderboard/requests/blob/main/01-ai/Yi-34B_eval_request_False_bfloat16_Original.json): it gives you information about the status of the evaluation*
-      - *The [aggregated results folder](https://huggingface.co/datasets/open-llm-leaderboard/results/tree/main/01-ai/Yi-34B): it gives you aggregated scores, per experimental run*
-      - *The [details dataset](https://huggingface.co/datasets/open-llm-leaderboard/details_01-ai__Yi-34B/tree/main): it gives you the full details (scores and examples for each task and a given model)*
-
-
-Why do models appear several times in the leaderboard? 
-- *We run evaluations with user-selected precision and model commit. Sometimes, users submit specific models at different commits and at different precisions (for example, in float16 and 4bit to see how quantization affects performance). You should be able to verify this by displaying the `precision` and `model sha` columns in the display. If, however, you see models appearing several times with the same precision and hash commit, this is not normal.*
-
-What is this concept of "flagging"?
-- *This mechanism allows users to report models that have unfair performance on the leaderboard. This contains several categories: exceedingly good results on the leaderboard because the model was (maybe accidentally) trained on the evaluation data, models that are copies of other models not attributed properly, etc.*
-
-My model has been flagged improperly, what can I do?
-- *Every flagged model has a discussion associated with it - feel free to plead your case there, and we'll see what to do together with the community.*
-
----------------------------
-
-## HOW TO SEARCH FOR A MODEL
-Search for models in the leaderboard by:
-1. Name, e.g., *model_name*
-2. Multiple names, separated by `;`, e.g., *model_name1;model_name2*
-3. License, prefix with `license:`, e.g., *license: MIT*
-4. Combination of name and license, order is irrelevant, e.g., *model_name; license: cc-by-sa-4.0*
-
----------------------------
-
-## EDITING SUBMISSIONS
-I upgraded my model and want to re-submit, how can I do that?
-- *Please open an issue with the precise name of your model, and we'll remove your model from the leaderboard so you can resubmit. You can also resubmit directly with the new commit hash!* 
-
-I need to rename my model, how can I do that?
-- *You can use @Weyaxi 's [super cool tool](https://huggingface.co/spaces/Weyaxi/open-llm-leaderboard-renamer) to request model name changes, then open a discussion where you link to the created pull request, and we'll check them and merge them as needed.*
-
----------------------------
-
-## OTHER
-Why do you differentiate between pretrained, continuously pretrained, fine-tuned, merges, etc?
-- *These different models do not play in the same categories, and therefore need to be separated for fair comparison. Base pretrained models are the most interesting for the community, as they are usually good models to fine-tune later on - any jump in performance from a pretrained model represents a true improvement on the SOTA. 
-Fine-tuned and IFT/RLHF/chat models usually have better performance, but the latter might be more sensitive to system prompts, which we do not cover at the moment in the Open LLM Leaderboard. 
-Merges and moerges have artificially inflated performance on test sets, which is not always explainable, and does not always apply to real-world situations.*
-
-What should I use the leaderboard for?
-- *We recommend using the leaderboard for 3 use cases: 1) getting an idea of the state of open pretrained models, by looking only at the ranks and score of this category; 2) experimenting with different fine-tuning methods, datasets, quantization techniques, etc, and comparing their score in a reproducible setup, and 3) checking the performance of a model of interest to you, wrt to other models of its category.*
-
-Why don't you display closed-source model scores? 
-- *This is a leaderboard for Open models, both for philosophical reasons (openness is cool) and for practical reasons: we want to ensure that the results we display are accurate and reproducible, but 1) commercial closed models can change their API thus rendering any scoring at a given time incorrect 2) we re-run everything on our cluster to ensure all models are run on the same setup and you can't do that for these models.*
-
-I have an issue with accessing the leaderboard through the Gradio API
-- *Since this is not the recommended way to access the leaderboard, we won't provide support for this, but you can look at tools provided by the community for inspiration!*
-
-I have another problem, help!
-- *Please open an issue in the discussion tab, and we'll do our best to help you in a timely manner :) *
-"""
-
-
-EVALUATION_QUEUE_TEXT = f"""
+EVALUATION_QUEUE_TEXT = """
 # Evaluation Queue for the 🤗 Open LLM Leaderboard
 
 Models added here will be automatically evaluated on the 🤗 cluster.
 
-## Don't forget to read the FAQ and the About tabs for more information!
-
-## First steps before submitting a model
+## Some good practices before submitting a model
 
 ### 1) Make sure you can load your model and tokenizer using AutoClasses:
 ```python
@@ -202,17 +93,16 @@ This is a leaderboard for Open LLMs, and we'd love for as many people as possibl
 ### 4) Fill up your model card
 When we add extra information about models to the leaderboard, it will be automatically taken from the model card
 
-### 5) Select the correct precision
-Not all models are converted properly from `float16` to `bfloat16`, and selecting the wrong precision can sometimes cause evaluation error (as loading a `bf16` model in `fp16` can sometimes generate NaNs, depending on the weight range).
-
-## Model types
-{icons}
+## 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"""
 @misc{open-llm-leaderboard,
-  author = {Edward Beeching and Clémentine Fourrier and Nathan Habib and Sheon Han and Nathan Lambert and Nazneen Rajani and Omar Sanseviero and Lewis Tunstall and Thomas Wolf},
+  author = {Edward Beeching, Clémentine Fourrier, Nathan Habib, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar Sanseviero, Lewis Tunstall, Thomas Wolf},
   title = {Open LLM Leaderboard},
   year = {2023},
   publisher = {Hugging Face},
@@ -275,32 +165,4 @@ CITATION_BUTTON_TEXT = r"""
       eprint={2109.07958},
       archivePrefix={arXiv},
       primaryClass={cs.CL}
-}
-@misc{DBLP:journals/corr/abs-1907-10641,
-      title={{WINOGRANDE:} An Adversarial Winograd Schema Challenge at Scale},
-      author={Keisuke Sakaguchi and Ronan Le Bras and Chandra Bhagavatula and Yejin Choi},
-      year={2019},
-      eprint={1907.10641},
-      archivePrefix={arXiv},
-      primaryClass={cs.CL}
-}
-@misc{DBLP:journals/corr/abs-2110-14168,
-      title={Training Verifiers to Solve Math Word Problems},
-      author={Karl Cobbe and
-                  Vineet Kosaraju and
-                  Mohammad Bavarian and
-                  Mark Chen and
-                  Heewoo Jun and
-                  Lukasz Kaiser and
-                  Matthias Plappert and
-                  Jerry Tworek and
-                  Jacob Hilton and
-                  Reiichiro Nakano and
-                  Christopher Hesse and
-                  John Schulman},
-      year={2021},
-      eprint={2110.14168},
-      archivePrefix={arXiv},
-      primaryClass={cs.CL}
-}
-"""
+}"""

src/display/formatting.py

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

src/display/utils.py

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

src/display_models/get_model_metadata.py

@@ -0,0 +1,167 @@
+import glob
+import json
+import os
+import re
+import pickle
+from typing import List
+
+import huggingface_hub
+from huggingface_hub import HfApi
+from tqdm import tqdm
+from transformers import AutoModel, AutoConfig
+from accelerate import init_empty_weights
+
+from src.display_models.model_metadata_flags import DO_NOT_SUBMIT_MODELS, FLAGGED_MODELS
+from src.display_models.model_metadata_type import MODEL_TYPE_METADATA, ModelType, model_type_from_str
+from src.display_models.utils import AutoEvalColumn, model_hyperlink
+
+api = HfApi(token=os.environ.get("H4_TOKEN", None))
+
+
+def get_model_infos_from_hub(leaderboard_data: List[dict]):
+    # load cache from disk
+    try:
+        with open("model_info_cache.pkl", "rb") as f:
+            model_info_cache = pickle.load(f)
+    except (EOFError, FileNotFoundError):
+        model_info_cache = {}
+    try:
+        with open("model_size_cache.pkl", "rb") as f:
+            model_size_cache = pickle.load(f)
+    except (EOFError, FileNotFoundError):
+        model_size_cache = {}
+
+    for model_data in tqdm(leaderboard_data):
+        model_name = model_data["model_name_for_query"]
+
+        if model_name in model_info_cache:
+            model_info = model_info_cache[model_name]
+        else:
+            try:
+                model_info = api.model_info(model_name)
+                model_info_cache[model_name] = model_info
+            except huggingface_hub.utils._errors.RepositoryNotFoundError:
+                print("Repo not found!", model_name)
+                model_data[AutoEvalColumn.license.name] = None
+                model_data[AutoEvalColumn.likes.name] = None
+                if model_name not in model_size_cache:
+                    model_size_cache[model_name] = get_model_size(model_name, None)
+                model_data[AutoEvalColumn.params.name] = model_size_cache[model_name]
+
+        model_data[AutoEvalColumn.license.name] = get_model_license(model_info)
+        model_data[AutoEvalColumn.likes.name] = get_model_likes(model_info)
+        if model_name not in model_size_cache:
+            model_size_cache[model_name] = get_model_size(model_name, model_info)
+        model_data[AutoEvalColumn.params.name] = model_size_cache[model_name]
+
+    # save cache to disk in pickle format
+    with open("model_info_cache.pkl", "wb") as f:
+        pickle.dump(model_info_cache, f)
+    with open("model_size_cache.pkl", "wb") as f:
+        pickle.dump(model_size_cache, f)
+
+
+def get_model_license(model_info):
+    try:
+        return model_info.cardData["license"]
+    except Exception:
+        return "?"
+
+
+def get_model_likes(model_info):
+    return model_info.likes
+
+
+size_pattern = re.compile(r"(\d\.)?\d+(b|m)")
+
+
+def get_model_size(model_name, model_info):
+    # In billions
+    try:
+        return round(model_info.safetensors["total"] / 1e9, 3)
+    except AttributeError:
+        try:
+            config = AutoConfig.from_pretrained(model_name, trust_remote_code=False)
+            with init_empty_weights():
+                model = AutoModel.from_config(config, trust_remote_code=False)
+            return round(sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e9, 3)
+        except (EnvironmentError, ValueError, KeyError):  # model config not found, likely private
+            try:
+                size_match = re.search(size_pattern, model_name.lower())
+                size = size_match.group(0)
+                return round(float(size[:-1]) if size[-1] == "b" else float(size[:-1]) / 1e3, 3)
+            except AttributeError:
+                return 0
+
+
+def get_model_type(leaderboard_data: List[dict]):
+    for model_data in leaderboard_data:
+        request_files = os.path.join(
+            "eval-queue",
+            model_data["model_name_for_query"] + "_eval_request_*" + ".json",
+        )
+        request_files = glob.glob(request_files)
+
+        # Select correct request file (precision)
+        request_file = ""
+        if len(request_files) == 1:
+            request_file = request_files[0]
+        elif len(request_files) > 1:
+            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"] == "FINISHED"
+                        and req_content["precision"] == model_data["Precision"].split(".")[-1]
+                    ):
+                        request_file = tmp_request_file
+
+        try:
+            with open(request_file, "r") as f:
+                request = json.load(f)
+            model_type = model_type_from_str(request["model_type"])
+            model_data[AutoEvalColumn.model_type.name] = model_type.value.name
+            model_data[AutoEvalColumn.model_type_symbol.name] = model_type.value.symbol  # + ("🔺" if is_delta else "")
+        except Exception:
+            if model_data["model_name_for_query"] in MODEL_TYPE_METADATA:
+                model_data[AutoEvalColumn.model_type.name] = MODEL_TYPE_METADATA[
+                    model_data["model_name_for_query"]
+                ].value.name
+                model_data[AutoEvalColumn.model_type_symbol.name] = MODEL_TYPE_METADATA[
+                    model_data["model_name_for_query"]
+                ].value.symbol  # + ("🔺" if is_delta else "")
+            else:
+                model_data[AutoEvalColumn.model_type.name] = ModelType.Unknown.value.name
+                model_data[AutoEvalColumn.model_type_symbol.name] = ModelType.Unknown.value.symbol
+
+
+def flag_models(leaderboard_data: List[dict]):
+    for model_data in leaderboard_data:
+        if model_data["model_name_for_query"] in FLAGGED_MODELS:
+            issue_num = FLAGGED_MODELS[model_data["model_name_for_query"]].split("/")[-1]
+            issue_link = model_hyperlink(
+                FLAGGED_MODELS[model_data["model_name_for_query"]],
+                f"See discussion #{issue_num}",
+            )
+            model_data[
+                AutoEvalColumn.model.name
+            ] = f"{model_data[AutoEvalColumn.model.name]} has been flagged! {issue_link}"
+
+
+def remove_forbidden_models(leaderboard_data: List[dict]):
+    indices_to_remove = []
+    for ix, model in enumerate(leaderboard_data):
+        if model["model_name_for_query"] in DO_NOT_SUBMIT_MODELS:
+            indices_to_remove.append(ix)
+
+    for ix in reversed(indices_to_remove):
+        leaderboard_data.pop(ix)
+    return leaderboard_data
+
+
+def apply_metadata(leaderboard_data: List[dict]):
+    leaderboard_data = remove_forbidden_models(leaderboard_data)
+    get_model_type(leaderboard_data)
+    get_model_infos_from_hub(leaderboard_data)
+    flag_models(leaderboard_data)

src/display_models/model_metadata_flags.py

@@ -0,0 +1,18 @@
+# Models which have been flagged by users as being problematic for a reason or another
+# (Model name to forum discussion link)
+FLAGGED_MODELS = {
+    "Voicelab/trurl-2-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/202",
+    "deepnight-research/llama-2-70B-inst": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/207",
+    "Aspik101/trurl-2-13b-pl-instruct_unload": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/213",
+    "Fredithefish/ReasonixPajama-3B-HF": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/236",
+    "TigerResearch/tigerbot-7b-sft-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/237",
+    "gaodrew/gaodrew-gorgonzola-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/215",
+    "AIDC-ai-business/Marcoroni-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+    "AIDC-ai-business/Marcoroni-13B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+    "AIDC-ai-business/Marcoroni-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
+}
+
+# Models which have been requested by orgs to not be submitted on the leaderboard
+DO_NOT_SUBMIT_MODELS = [
+    "Voicelab/trurl-2-13b",  # trained on MMLU
+]

src/display_models/model_metadata_type.py

@@ -0,0 +1,555 @@
+from dataclasses import dataclass
+from enum import Enum
+from typing import Dict
+
+
+@dataclass
+class ModelInfo:
+    name: str
+    symbol: str  # emoji
+
+
+class ModelType(Enum):
+    PT = ModelInfo(name="pretrained", symbol="🟢")
+    FT = ModelInfo(name="fine-tuned", symbol="🔶")
+    IFT = ModelInfo(name="instruction-tuned", symbol="⭕")
+    RL = ModelInfo(name="RL-tuned", symbol="🟦")
+    Unknown = ModelInfo(name="Unknown", symbol="?")
+
+    def to_str(self, separator=" "):
+        return f"{self.value.symbol}{separator}{self.value.name}"
+
+
+MODEL_TYPE_METADATA: Dict[str, ModelType] = {
+    "tiiuae/falcon-180B": ModelType.PT,
+    "tiiuae/falcon-180B-chat": ModelType.RL,
+    "microsoft/phi-1_5": ModelType.PT,
+    "Qwen/Qwen-7B": ModelType.PT,
+    "Qwen/Qwen-7B-Chat": ModelType.RL,
+    "notstoic/PygmalionCoT-7b": ModelType.IFT,
+    "aisquared/dlite-v1-355m": ModelType.IFT,
+    "aisquared/dlite-v1-1_5b": ModelType.IFT,
+    "aisquared/dlite-v1-774m": ModelType.IFT,
+    "aisquared/dlite-v1-124m": ModelType.IFT,
+    "aisquared/chopt-2_7b": ModelType.IFT,
+    "aisquared/dlite-v2-124m": ModelType.IFT,
+    "aisquared/dlite-v2-774m": ModelType.IFT,
+    "aisquared/dlite-v2-1_5b": ModelType.IFT,
+    "aisquared/chopt-1_3b": ModelType.IFT,
+    "aisquared/dlite-v2-355m": ModelType.IFT,
+    "augtoma/qCammel-13": ModelType.IFT,
+    "Aspik101/Llama-2-7b-hf-instruct-pl-lora_unload": ModelType.IFT,
+    "Aspik101/vicuna-7b-v1.3-instruct-pl-lora_unload": ModelType.IFT,
+    "TheBloke/alpaca-lora-65B-HF": ModelType.FT,
+    "TheBloke/tulu-7B-fp16": ModelType.IFT,
+    "TheBloke/guanaco-7B-HF": ModelType.FT,
+    "TheBloke/koala-7B-HF": ModelType.FT,
+    "TheBloke/wizardLM-7B-HF": ModelType.IFT,
+    "TheBloke/airoboros-13B-HF": ModelType.IFT,
+    "TheBloke/koala-13B-HF": ModelType.FT,
+    "TheBloke/Wizard-Vicuna-7B-Uncensored-HF": ModelType.FT,
+    "TheBloke/dromedary-65b-lora-HF": ModelType.IFT,
+    "TheBloke/wizardLM-13B-1.0-fp16": ModelType.IFT,
+    "TheBloke/WizardLM-13B-V1-1-SuperHOT-8K-fp16": ModelType.FT,
+    "TheBloke/Wizard-Vicuna-30B-Uncensored-fp16": ModelType.FT,
+    "TheBloke/wizard-vicuna-13B-HF": ModelType.IFT,
+    "TheBloke/UltraLM-13B-fp16": ModelType.IFT,
+    "TheBloke/OpenAssistant-FT-7-Llama-30B-HF": ModelType.FT,
+    "TheBloke/vicuna-13B-1.1-HF": ModelType.IFT,
+    "TheBloke/guanaco-13B-HF": ModelType.FT,
+    "TheBloke/guanaco-65B-HF": ModelType.FT,
+    "TheBloke/airoboros-7b-gpt4-fp16": ModelType.IFT,
+    "TheBloke/llama-30b-supercot-SuperHOT-8K-fp16": ModelType.IFT,
+    "TheBloke/Llama-2-13B-fp16": ModelType.PT,
+    "TheBloke/llama-2-70b-Guanaco-QLoRA-fp16": ModelType.FT,
+    "TheBloke/landmark-attention-llama7b-fp16": ModelType.IFT,
+    "TheBloke/Planner-7B-fp16": ModelType.IFT,
+    "TheBloke/Wizard-Vicuna-13B-Uncensored-HF": ModelType.FT,
+    "TheBloke/gpt4-alpaca-lora-13B-HF": ModelType.IFT,
+    "TheBloke/gpt4-x-vicuna-13B-HF": ModelType.IFT,
+    "TheBloke/gpt4-alpaca-lora_mlp-65B-HF": ModelType.IFT,
+    "TheBloke/tulu-13B-fp16": ModelType.IFT,
+    "TheBloke/VicUnlocked-alpaca-65B-QLoRA-fp16": ModelType.IFT,
+    "TheBloke/Llama-2-70B-fp16": ModelType.IFT,
+    "TheBloke/WizardLM-30B-fp16": ModelType.IFT,
+    "TheBloke/robin-13B-v2-fp16": ModelType.FT,
+    "TheBloke/robin-33B-v2-fp16": ModelType.FT,
+    "TheBloke/Vicuna-13B-CoT-fp16": ModelType.IFT,
+    "TheBloke/Vicuna-33B-1-3-SuperHOT-8K-fp16": ModelType.IFT,
+    "TheBloke/Wizard-Vicuna-30B-Superhot-8K-fp16": ModelType.FT,
+    "TheBloke/Nous-Hermes-13B-SuperHOT-8K-fp16": ModelType.IFT,
+    "TheBloke/GPlatty-30B-SuperHOT-8K-fp16": ModelType.FT,
+    "TheBloke/CAMEL-33B-Combined-Data-SuperHOT-8K-fp16": ModelType.IFT,
+    "TheBloke/Chinese-Alpaca-33B-SuperHOT-8K-fp16": ModelType.IFT,
+    "jphme/orca_mini_v2_ger_7b": ModelType.IFT,
+    "Ejafa/vicuna_7B_vanilla_1.1": ModelType.FT,
+    "kevinpro/Vicuna-13B-CoT": ModelType.IFT,
+    "AlekseyKorshuk/pygmalion-6b-vicuna-chatml": ModelType.FT,
+    "AlekseyKorshuk/chatml-pyg-v1": ModelType.FT,
+    "concedo/Vicuzard-30B-Uncensored": ModelType.FT,
+    "concedo/OPT-19M-ChatSalad": ModelType.FT,
+    "concedo/Pythia-70M-ChatSalad": ModelType.FT,
+    "digitous/13B-HyperMantis": ModelType.IFT,
+    "digitous/Adventien-GPTJ": ModelType.FT,
+    "digitous/Alpacino13b": ModelType.IFT,
+    "digitous/GPT-R": ModelType.IFT,
+    "digitous/Javelin-R": ModelType.IFT,
+    "digitous/Javalion-GPTJ": ModelType.IFT,
+    "digitous/Javalion-R": ModelType.IFT,
+    "digitous/Skegma-GPTJ": ModelType.FT,
+    "digitous/Alpacino30b": ModelType.IFT,
+    "digitous/Janin-GPTJ": ModelType.FT,
+    "digitous/Janin-R": ModelType.FT,
+    "digitous/Javelin-GPTJ": ModelType.FT,
+    "SaylorTwift/gpt2_test": ModelType.PT,
+    "anton-l/gpt-j-tiny-random": ModelType.FT,
+    "Andron00e/YetAnother_Open-Llama-3B-LoRA-OpenOrca": ModelType.FT,
+    "Lazycuber/pyg-instruct-wizardlm": ModelType.FT,
+    "Lazycuber/Janemalion-6B": ModelType.FT,
+    "IDEA-CCNL/Ziya-LLaMA-13B-Pretrain-v1": ModelType.FT,
+    "IDEA-CCNL/Ziya-LLaMA-13B-v1": ModelType.IFT,
+    "dsvv-cair/alpaca-cleaned-llama-30b-bf16": ModelType.FT,
+    "gpt2-medium": ModelType.PT,
+    "camel-ai/CAMEL-13B-Combined-Data": ModelType.IFT,
+    "camel-ai/CAMEL-13B-Role-Playing-Data": ModelType.FT,
+    "camel-ai/CAMEL-33B-Combined-Data": ModelType.IFT,
+    "PygmalionAI/pygmalion-6b": ModelType.FT,
+    "PygmalionAI/metharme-1.3b": ModelType.IFT,
+    "PygmalionAI/pygmalion-1.3b": ModelType.FT,
+    "PygmalionAI/pygmalion-350m": ModelType.FT,
+    "PygmalionAI/pygmalion-2.7b": ModelType.FT,
+    "medalpaca/medalpaca-7b": ModelType.FT,
+    "lilloukas/Platypus-30B": ModelType.IFT,
+    "lilloukas/GPlatty-30B": ModelType.FT,
+    "mncai/chatdoctor": ModelType.FT,
+    "chaoyi-wu/MedLLaMA_13B": ModelType.FT,
+    "LoupGarou/WizardCoder-Guanaco-15B-V1.0": ModelType.IFT,
+    "LoupGarou/WizardCoder-Guanaco-15B-V1.1": ModelType.FT,
+    "hakurei/instruct-12b": ModelType.IFT,
+    "hakurei/lotus-12B": ModelType.FT,
+    "shibing624/chinese-llama-plus-13b-hf": ModelType.IFT,
+    "shibing624/chinese-alpaca-plus-7b-hf": ModelType.IFT,
+    "shibing624/chinese-alpaca-plus-13b-hf": ModelType.IFT,
+    "mosaicml/mpt-7b-instruct": ModelType.IFT,
+    "mosaicml/mpt-30b-chat": ModelType.IFT,
+    "mosaicml/mpt-7b-storywriter": ModelType.FT,
+    "mosaicml/mpt-30b-instruct": ModelType.IFT,
+    "mosaicml/mpt-7b-chat": ModelType.IFT,
+    "mosaicml/mpt-30b": ModelType.PT,
+    "Corianas/111m": ModelType.IFT,
+    "Corianas/Quokka_1.3b": ModelType.IFT,
+    "Corianas/256_5epoch": ModelType.FT,
+    "Corianas/Quokka_256m": ModelType.IFT,
+    "Corianas/Quokka_590m": ModelType.IFT,
+    "Corianas/gpt-j-6B-Dolly": ModelType.FT,
+    "Corianas/Quokka_2.7b": ModelType.IFT,
+    "cyberagent/open-calm-7b": ModelType.FT,
+    "Aspik101/Nous-Hermes-13b-pl-lora_unload": ModelType.IFT,
+    "THUDM/chatglm2-6b": ModelType.IFT,
+    "MetaIX/GPT4-X-Alpasta-30b": ModelType.IFT,
+    "NYTK/PULI-GPTrio": ModelType.PT,
+    "EleutherAI/pythia-1.3b": ModelType.PT,
+    "EleutherAI/pythia-2.8b-deduped": ModelType.PT,
+    "EleutherAI/gpt-neo-125m": ModelType.PT,
+    "EleutherAI/pythia-160m": ModelType.PT,
+    "EleutherAI/gpt-neo-2.7B": ModelType.PT,
+    "EleutherAI/pythia-1b-deduped": ModelType.PT,
+    "EleutherAI/pythia-6.7b": ModelType.PT,
+    "EleutherAI/pythia-70m-deduped": ModelType.PT,
+    "EleutherAI/gpt-neox-20b": ModelType.PT,
+    "EleutherAI/pythia-1.4b-deduped": ModelType.PT,
+    "EleutherAI/pythia-2.7b": ModelType.PT,
+    "EleutherAI/pythia-6.9b-deduped": ModelType.PT,
+    "EleutherAI/pythia-70m": ModelType.PT,
+    "EleutherAI/gpt-j-6b": ModelType.PT,
+    "EleutherAI/pythia-12b-deduped": ModelType.PT,
+    "EleutherAI/gpt-neo-1.3B": ModelType.PT,
+    "EleutherAI/pythia-410m-deduped": ModelType.PT,
+    "EleutherAI/pythia-160m-deduped": ModelType.PT,
+    "EleutherAI/polyglot-ko-12.8b": ModelType.PT,
+    "EleutherAI/pythia-12b": ModelType.PT,
+    "roneneldan/TinyStories-33M": ModelType.PT,
+    "roneneldan/TinyStories-28M": ModelType.PT,
+    "roneneldan/TinyStories-1M": ModelType.PT,
+    "roneneldan/TinyStories-8M": ModelType.PT,
+    "roneneldan/TinyStories-3M": ModelType.PT,
+    "jerryjalapeno/nart-100k-7b": ModelType.FT,
+    "lmsys/vicuna-13b-v1.3": ModelType.IFT,
+    "lmsys/vicuna-7b-v1.3": ModelType.IFT,
+    "lmsys/vicuna-13b-v1.1": ModelType.IFT,
+    "lmsys/vicuna-13b-delta-v1.1": ModelType.IFT,
+    "lmsys/vicuna-7b-delta-v1.1": ModelType.IFT,
+    "abhiramtirumala/DialoGPT-sarcastic-medium": ModelType.FT,
+    "haonan-li/bactrian-x-llama-13b-merged": ModelType.IFT,
+    "Gryphe/MythoLogic-13b": ModelType.IFT,
+    "Gryphe/MythoBoros-13b": ModelType.IFT,
+    "pillowtalks-ai/delta13b": ModelType.FT,
+    "wannaphong/openthaigpt-0.1.0-beta-full-model_for_open_llm_leaderboard": ModelType.FT,
+    "bigscience/bloom-7b1": ModelType.PT,
+    "bigcode/tiny_starcoder_py": ModelType.PT,
+    "bigcode/starcoderplus": ModelType.FT,
+    "bigcode/gpt_bigcode-santacoder": ModelType.PT,
+    "bigcode/starcoder": ModelType.PT,
+    "Open-Orca/OpenOrca-Preview1-13B": ModelType.IFT,
+    "microsoft/DialoGPT-large": ModelType.FT,
+    "microsoft/DialoGPT-small": ModelType.FT,
+    "microsoft/DialoGPT-medium": ModelType.FT,
+    "microsoft/CodeGPT-small-py": ModelType.FT,
+    "Tincando/fiction_story_generator": ModelType.FT,
+    "Pirr/pythia-13b-deduped-green_devil": ModelType.FT,
+    "Aeala/GPT4-x-AlpacaDente2-30b": ModelType.FT,
+    "Aeala/GPT4-x-AlpacaDente-30b": ModelType.FT,
+    "Aeala/GPT4-x-Alpasta-13b": ModelType.FT,
+    "Aeala/VicUnlocked-alpaca-30b": ModelType.IFT,
+    "Tap-M/Luna-AI-Llama2-Uncensored": ModelType.FT,
+    "illuin/test-custom-llama": ModelType.FT,
+    "dvruette/oasst-llama-13b-2-epochs": ModelType.FT,
+    "dvruette/oasst-gpt-neox-20b-1000-steps": ModelType.FT,
+    "dvruette/llama-13b-pretrained-dropout": ModelType.PT,
+    "dvruette/llama-13b-pretrained": ModelType.PT,
+    "dvruette/llama-13b-pretrained-sft-epoch-1": ModelType.FT,
+    "dvruette/llama-13b-pretrained-sft-do2": ModelType.FT,
+    "dvruette/oasst-gpt-neox-20b-3000-steps": ModelType.FT,
+    "dvruette/oasst-pythia-12b-pretrained-sft": ModelType.FT,
+    "dvruette/oasst-pythia-6.9b-4000-steps": ModelType.FT,
+    "dvruette/gpt-neox-20b-full-precision": ModelType.FT,
+    "dvruette/oasst-llama-13b-1000-steps": ModelType.FT,
+    "openlm-research/open_llama_7b_700bt_preview": ModelType.PT,
+    "openlm-research/open_llama_7b": ModelType.PT,
+    "openlm-research/open_llama_7b_v2": ModelType.PT,
+    "openlm-research/open_llama_3b": ModelType.PT,
+    "openlm-research/open_llama_13b": ModelType.PT,
+    "openlm-research/open_llama_3b_v2": ModelType.PT,
+    "PocketDoc/Dans-PileOfSets-Mk1-llama-13b-merged": ModelType.IFT,
+    "GeorgiaTechResearchInstitute/galpaca-30b": ModelType.IFT,
+    "GeorgiaTechResearchInstitute/starcoder-gpteacher-code-instruct": ModelType.IFT,
+    "databricks/dolly-v2-7b": ModelType.IFT,
+    "databricks/dolly-v2-3b": ModelType.IFT,
+    "databricks/dolly-v2-12b": ModelType.IFT,
+    "Rachneet/gpt2-xl-alpaca": ModelType.FT,
+    "Locutusque/gpt2-conversational-or-qa": ModelType.FT,
+    "psyche/kogpt": ModelType.FT,
+    "NbAiLab/nb-gpt-j-6B-alpaca": ModelType.IFT,
+    "Mikael110/llama-2-7b-guanaco-fp16": ModelType.FT,
+    "Mikael110/llama-2-13b-guanaco-fp16": ModelType.FT,
+    "Fredithefish/CrimsonPajama": ModelType.IFT,
+    "Fredithefish/RedPajama-INCITE-Chat-3B-ShareGPT-11K": ModelType.FT,
+    "Fredithefish/ScarletPajama-3B-HF": ModelType.FT,
+    "Fredithefish/RedPajama-INCITE-Chat-3B-Instruction-Tuning-with-GPT-4": ModelType.IFT,
+    "acrastt/RedPajama-INCITE-Chat-Instruct-3B-V1": ModelType.IFT,
+    "eachadea/vicuna-13b-1.1": ModelType.FT,
+    "eachadea/vicuna-7b-1.1": ModelType.FT,
+    "eachadea/vicuna-13b": ModelType.FT,
+    "openaccess-ai-collective/wizard-mega-13b": ModelType.IFT,
+    "openaccess-ai-collective/manticore-13b": ModelType.IFT,
+    "openaccess-ai-collective/manticore-30b-chat-pyg-alpha": ModelType.IFT,
+    "openaccess-ai-collective/minotaur-13b": ModelType.IFT,
+    "openaccess-ai-collective/minotaur-13b-fixed": ModelType.IFT,
+    "openaccess-ai-collective/hippogriff-30b-chat": ModelType.IFT,
+    "openaccess-ai-collective/manticore-13b-chat-pyg": ModelType.IFT,
+    "pythainlp/wangchanglm-7.5B-sft-enth": ModelType.IFT,
+    "pythainlp/wangchanglm-7.5B-sft-en-sharded": ModelType.IFT,
+    "euclaise/gpt-neox-122m-minipile-digits": ModelType.FT,
+    "stabilityai/StableBeluga1-Delta": ModelType.IFT,
+    "stabilityai/stablelm-tuned-alpha-7b": ModelType.IFT,
+    "stabilityai/StableBeluga2": ModelType.IFT,
+    "stabilityai/StableBeluga-13B": ModelType.IFT,
+    "stabilityai/StableBeluga-7B": ModelType.IFT,
+    "stabilityai/stablelm-base-alpha-7b": ModelType.PT,
+    "stabilityai/stablelm-base-alpha-3b": ModelType.PT,
+    "stabilityai/stablelm-tuned-alpha-3b": ModelType.IFT,
+    "alibidaran/medical_transcription_generator": ModelType.FT,
+    "CalderaAI/30B-Lazarus": ModelType.IFT,
+    "CalderaAI/13B-BlueMethod": ModelType.IFT,
+    "CalderaAI/13B-Ouroboros": ModelType.IFT,
+    "KoboldAI/OPT-13B-Erebus": ModelType.FT,
+    "KoboldAI/GPT-J-6B-Janeway": ModelType.FT,
+    "KoboldAI/GPT-J-6B-Shinen": ModelType.FT,
+    "KoboldAI/fairseq-dense-2.7B": ModelType.PT,
+    "KoboldAI/OPT-6B-nerys-v2": ModelType.FT,
+    "KoboldAI/GPT-NeoX-20B-Skein": ModelType.FT,
+    "KoboldAI/PPO_Pygway-6b-Mix": ModelType.FT,
+    "KoboldAI/fairseq-dense-6.7B": ModelType.PT,
+    "KoboldAI/fairseq-dense-125M": ModelType.PT,
+    "KoboldAI/OPT-13B-Nerybus-Mix": ModelType.FT,
+    "KoboldAI/OPT-2.7B-Erebus": ModelType.FT,
+    "KoboldAI/OPT-350M-Nerys-v2": ModelType.FT,
+    "KoboldAI/OPT-2.7B-Nerys-v2": ModelType.FT,
+    "KoboldAI/OPT-2.7B-Nerybus-Mix": ModelType.FT,
+    "KoboldAI/OPT-13B-Nerys-v2": ModelType.FT,
+    "KoboldAI/GPT-NeoX-20B-Erebus": ModelType.FT,
+    "KoboldAI/OPT-6.7B-Erebus": ModelType.FT,
+    "KoboldAI/fairseq-dense-355M": ModelType.PT,
+    "KoboldAI/OPT-6.7B-Nerybus-Mix": ModelType.FT,
+    "KoboldAI/GPT-J-6B-Adventure": ModelType.FT,
+    "KoboldAI/OPT-350M-Erebus": ModelType.FT,
+    "KoboldAI/GPT-J-6B-Skein": ModelType.FT,
+    "KoboldAI/OPT-30B-Erebus": ModelType.FT,
+    "klosax/pythia-160m-deduped-step92k-193bt": ModelType.PT,
+    "klosax/open_llama_3b_350bt_preview": ModelType.PT,
+    "klosax/openllama-3b-350bt": ModelType.PT,
+    "klosax/pythia-70m-deduped-step44k-92bt": ModelType.PT,
+    "klosax/open_llama_13b_600bt_preview": ModelType.PT,
+    "klosax/open_llama_7b_400bt_preview": ModelType.PT,
+    "kfkas/Llama-2-ko-7b-Chat": ModelType.IFT,
+    "WeOpenML/Alpaca-7B-v1": ModelType.IFT,
+    "WeOpenML/PandaLM-Alpaca-7B-v1": ModelType.IFT,
+    "TFLai/gpt2-turkish-uncased": ModelType.FT,
+    "ehartford/WizardLM-13B-Uncensored": ModelType.IFT,
+    "ehartford/dolphin-llama-13b": ModelType.IFT,
+    "ehartford/Wizard-Vicuna-30B-Uncensored": ModelType.FT,
+    "ehartford/WizardLM-30B-Uncensored": ModelType.IFT,
+    "ehartford/Wizard-Vicuna-13B-Uncensored": ModelType.FT,
+    "ehartford/WizardLM-7B-Uncensored": ModelType.IFT,
+    "ehartford/based-30b": ModelType.FT,
+    "ehartford/Wizard-Vicuna-7B-Uncensored": ModelType.FT,
+    "wahaha1987/llama_7b_sharegpt94k_fastchat": ModelType.FT,
+    "wahaha1987/llama_13b_sharegpt94k_fastchat": ModelType.FT,
+    "OpenAssistant/oasst-sft-1-pythia-12b": ModelType.FT,
+    "OpenAssistant/stablelm-7b-sft-v7-epoch-3": ModelType.IFT,
+    "OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5": ModelType.FT,
+    "OpenAssistant/pythia-12b-sft-v8-2.5k-steps": ModelType.IFT,
+    "OpenAssistant/pythia-12b-sft-v8-7k-steps": ModelType.IFT,
+    "OpenAssistant/pythia-12b-pre-v8-12.5k-steps": ModelType.IFT,
+    "OpenAssistant/llama2-13b-orca-8k-3319": ModelType.IFT,
+    "junelee/wizard-vicuna-13b": ModelType.FT,
+    "BreadAi/gpt-YA-1-1_160M": ModelType.PT,
+    "BreadAi/MuseCan": ModelType.PT,
+    "BreadAi/MusePy-1-2": ModelType.PT,
+    "BreadAi/DiscordPy": ModelType.PT,
+    "BreadAi/PM_modelV2": ModelType.PT,
+    "BreadAi/gpt-Youtube": ModelType.PT,
+    "BreadAi/StoryPy": ModelType.FT,
+    "julianweng/Llama-2-7b-chat-orcah": ModelType.FT,
+    "AGI-inc/lora_moe_7b_baseline": ModelType.FT,
+    "AGI-inc/lora_moe_7b": ModelType.FT,
+    "togethercomputer/GPT-NeoXT-Chat-Base-20B": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Chat-7B-v0.1": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Instruct-7B-v0.1": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-7B-Base": ModelType.PT,
+    "togethercomputer/RedPajama-INCITE-7B-Instruct": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Base-3B-v1": ModelType.PT,
+    "togethercomputer/Pythia-Chat-Base-7B": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Base-7B-v0.1": ModelType.PT,
+    "togethercomputer/GPT-JT-6B-v1": ModelType.IFT,
+    "togethercomputer/GPT-JT-6B-v0": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Chat-3B-v1": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-7B-Chat": ModelType.IFT,
+    "togethercomputer/RedPajama-INCITE-Instruct-3B-v1": ModelType.IFT,
+    "Writer/camel-5b-hf": ModelType.IFT,
+    "Writer/palmyra-base": ModelType.PT,
+    "MBZUAI/LaMini-GPT-1.5B": ModelType.IFT,
+    "MBZUAI/lamini-cerebras-111m": ModelType.IFT,
+    "MBZUAI/lamini-neo-1.3b": ModelType.IFT,
+    "MBZUAI/lamini-cerebras-1.3b": ModelType.IFT,
+    "MBZUAI/lamini-cerebras-256m": ModelType.IFT,
+    "MBZUAI/LaMini-GPT-124M": ModelType.IFT,
+    "MBZUAI/lamini-neo-125m": ModelType.IFT,
+    "TehVenom/DiffMerge-DollyGPT-Pygmalion": ModelType.FT,
+    "TehVenom/PPO_Shygmalion-6b": ModelType.FT,
+    "TehVenom/Dolly_Shygmalion-6b-Dev_V8P2": ModelType.FT,
+    "TehVenom/Pygmalion_AlpacaLora-7b": ModelType.FT,
+    "TehVenom/PPO_Pygway-V8p4_Dev-6b": ModelType.FT,
+    "TehVenom/Dolly_Malion-6b": ModelType.FT,
+    "TehVenom/PPO_Shygmalion-V8p4_Dev-6b": ModelType.FT,
+    "TehVenom/ChanMalion": ModelType.FT,
+    "TehVenom/GPT-J-Pyg_PPO-6B": ModelType.IFT,
+    "TehVenom/Pygmalion-13b-Merged": ModelType.FT,
+    "TehVenom/Metharme-13b-Merged": ModelType.IFT,
+    "TehVenom/Dolly_Shygmalion-6b": ModelType.FT,
+    "TehVenom/GPT-J-Pyg_PPO-6B-Dev-V8p4": ModelType.IFT,
+    "georgesung/llama2_7b_chat_uncensored": ModelType.FT,
+    "vicgalle/gpt2-alpaca": ModelType.IFT,
+    "vicgalle/alpaca-7b": ModelType.FT,
+    "vicgalle/gpt2-alpaca-gpt4": ModelType.IFT,
+    "facebook/opt-350m": ModelType.PT,
+    "facebook/opt-125m": ModelType.PT,
+    "facebook/xglm-4.5B": ModelType.PT,
+    "facebook/opt-2.7b": ModelType.PT,
+    "facebook/opt-6.7b": ModelType.PT,
+    "facebook/galactica-30b": ModelType.PT,
+    "facebook/opt-13b": ModelType.PT,
+    "facebook/opt-66b": ModelType.PT,
+    "facebook/xglm-7.5B": ModelType.PT,
+    "facebook/xglm-564M": ModelType.PT,
+    "facebook/opt-30b": ModelType.PT,
+    "golaxy/gogpt-7b": ModelType.FT,
+    "golaxy/gogpt2-7b": ModelType.FT,
+    "golaxy/gogpt-7b-bloom": ModelType.FT,
+    "golaxy/gogpt-3b-bloom": ModelType.FT,
+    "psmathur/orca_mini_v2_7b": ModelType.IFT,
+    "psmathur/orca_mini_7b": ModelType.IFT,
+    "psmathur/orca_mini_3b": ModelType.IFT,
+    "psmathur/orca_mini_v2_13b": ModelType.IFT,
+    "gpt2-xl": ModelType.PT,
+    "lxe/Cerebras-GPT-2.7B-Alpaca-SP": ModelType.FT,
+    "Monero/Manticore-13b-Chat-Pyg-Guanaco": ModelType.FT,
+    "Monero/WizardLM-Uncensored-SuperCOT-StoryTelling-30b": ModelType.IFT,
+    "Monero/WizardLM-13b-OpenAssistant-Uncensored": ModelType.IFT,
+    "Monero/WizardLM-30B-Uncensored-Guanaco-SuperCOT-30b": ModelType.IFT,
+    "jzjiao/opt-1.3b-rlhf": ModelType.FT,
+    "HuggingFaceH4/starchat-beta": ModelType.IFT,
+    "KnutJaegersberg/gpt-2-xl-EvolInstruct": ModelType.IFT,
+    "KnutJaegersberg/megatron-GPT-2-345m-EvolInstruct": ModelType.IFT,
+    "KnutJaegersberg/galactica-orca-wizardlm-1.3b": ModelType.IFT,
+    "openchat/openchat_8192": ModelType.IFT,
+    "openchat/openchat_v2": ModelType.IFT,
+    "openchat/openchat_v2_w": ModelType.IFT,
+    "ausboss/llama-13b-supercot": ModelType.IFT,
+    "ausboss/llama-30b-supercot": ModelType.IFT,
+    "Neko-Institute-of-Science/metharme-7b": ModelType.IFT,
+    "Neko-Institute-of-Science/pygmalion-7b": ModelType.FT,
+    "SebastianSchramm/Cerebras-GPT-111M-instruction": ModelType.IFT,
+    "victor123/WizardLM-13B-1.0": ModelType.IFT,
+    "OpenBuddy/openbuddy-openllama-13b-v7-fp16": ModelType.FT,
+    "OpenBuddy/openbuddy-llama2-13b-v8.1-fp16": ModelType.FT,
+    "OpenBuddyEA/openbuddy-llama-30b-v7.1-bf16": ModelType.FT,
+    "baichuan-inc/Baichuan-7B": ModelType.PT,
+    "tiiuae/falcon-40b-instruct": ModelType.IFT,
+    "tiiuae/falcon-40b": ModelType.PT,
+    "tiiuae/falcon-7b": ModelType.PT,
+    "YeungNLP/firefly-llama-13b": ModelType.FT,
+    "YeungNLP/firefly-llama-13b-v1.2": ModelType.FT,
+    "YeungNLP/firefly-llama2-13b": ModelType.FT,
+    "YeungNLP/firefly-ziya-13b": ModelType.FT,
+    "shaohang/Sparse0.5_OPT-1.3": ModelType.FT,
+    "xzuyn/Alpacino-SuperCOT-13B": ModelType.IFT,
+    "xzuyn/MedicWizard-7B": ModelType.FT,
+    "xDAN-AI/xDAN_13b_l2_lora": ModelType.FT,
+    "beomi/KoAlpaca-Polyglot-5.8B": ModelType.FT,
+    "beomi/llama-2-ko-7b": ModelType.IFT,
+    "Salesforce/codegen-6B-multi": ModelType.PT,
+    "Salesforce/codegen-16B-nl": ModelType.PT,
+    "Salesforce/codegen-6B-nl": ModelType.PT,
+    "ai-forever/rugpt3large_based_on_gpt2": ModelType.FT,
+    "gpt2-large": ModelType.PT,
+    "frank098/orca_mini_3b_juniper": ModelType.FT,
+    "frank098/WizardLM_13B_juniper": ModelType.FT,
+    "FPHam/Free_Sydney_13b_HF": ModelType.FT,
+    "huggingface/llama-13b": ModelType.PT,
+    "huggingface/llama-7b": ModelType.PT,
+    "huggingface/llama-65b": ModelType.PT,
+    "huggingface/llama-30b": ModelType.PT,
+    "Henk717/chronoboros-33B": ModelType.IFT,
+    "jondurbin/airoboros-13b-gpt4-1.4": ModelType.IFT,
+    "jondurbin/airoboros-7b": ModelType.IFT,
+    "jondurbin/airoboros-7b-gpt4": ModelType.IFT,
+    "jondurbin/airoboros-7b-gpt4-1.1": ModelType.IFT,
+    "jondurbin/airoboros-7b-gpt4-1.2": ModelType.IFT,
+    "jondurbin/airoboros-7b-gpt4-1.3": ModelType.IFT,
+    "jondurbin/airoboros-7b-gpt4-1.4": ModelType.IFT,
+    "jondurbin/airoboros-l2-7b-gpt4-1.4.1": ModelType.IFT,
+    "jondurbin/airoboros-l2-13b-gpt4-1.4.1": ModelType.IFT,
+    "jondurbin/airoboros-l2-70b-gpt4-1.4.1": ModelType.IFT,
+    "jondurbin/airoboros-13b": ModelType.IFT,
+    "jondurbin/airoboros-33b-gpt4-1.4": ModelType.IFT,
+    "jondurbin/airoboros-33b-gpt4-1.2": ModelType.IFT,
+    "jondurbin/airoboros-65b-gpt4-1.2": ModelType.IFT,
+    "ariellee/SuperPlatty-30B": ModelType.IFT,
+    "danielhanchen/open_llama_3b_600bt_preview": ModelType.FT,
+    "cerebras/Cerebras-GPT-256M": ModelType.PT,
+    "cerebras/Cerebras-GPT-1.3B": ModelType.PT,
+    "cerebras/Cerebras-GPT-13B": ModelType.PT,
+    "cerebras/Cerebras-GPT-2.7B": ModelType.PT,
+    "cerebras/Cerebras-GPT-111M": ModelType.PT,
+    "cerebras/Cerebras-GPT-6.7B": ModelType.PT,
+    "Yhyu13/oasst-rlhf-2-llama-30b-7k-steps-hf": ModelType.RL,
+    "Yhyu13/llama-30B-hf-openassitant": ModelType.FT,
+    "NousResearch/Nous-Hermes-Llama2-13b": ModelType.IFT,
+    "NousResearch/Nous-Hermes-llama-2-7b": ModelType.IFT,
+    "NousResearch/Redmond-Puffin-13B": ModelType.IFT,
+    "NousResearch/Nous-Hermes-13b": ModelType.IFT,
+    "project-baize/baize-v2-7b": ModelType.IFT,
+    "project-baize/baize-v2-13b": ModelType.IFT,
+    "LLMs/WizardLM-13B-V1.0": ModelType.FT,
+    "LLMs/AlpacaGPT4-7B-elina": ModelType.FT,
+    "wenge-research/yayi-7b": ModelType.FT,
+    "wenge-research/yayi-7b-llama2": ModelType.FT,
+    "wenge-research/yayi-13b-llama2": ModelType.FT,
+    "yhyhy3/open_llama_7b_v2_med_instruct": ModelType.IFT,
+    "llama-anon/instruct-13b": ModelType.IFT,
+    "huggingtweets/jerma985": ModelType.FT,
+    "huggingtweets/gladosystem": ModelType.FT,
+    "huggingtweets/bladeecity-jerma985": ModelType.FT,
+    "huggyllama/llama-13b": ModelType.PT,
+    "huggyllama/llama-65b": ModelType.PT,
+    "FabbriSimo01/Facebook_opt_1.3b_Quantized": ModelType.PT,
+    "upstage/Llama-2-70b-instruct": ModelType.IFT,
+    "upstage/Llama-2-70b-instruct-1024": ModelType.IFT,
+    "upstage/llama-65b-instruct": ModelType.IFT,
+    "upstage/llama-30b-instruct-2048": ModelType.IFT,
+    "upstage/llama-30b-instruct": ModelType.IFT,
+    "WizardLM/WizardLM-13B-1.0": ModelType.IFT,
+    "WizardLM/WizardLM-13B-V1.1": ModelType.IFT,
+    "WizardLM/WizardLM-13B-V1.2": ModelType.IFT,
+    "WizardLM/WizardLM-30B-V1.0": ModelType.IFT,
+    "WizardLM/WizardCoder-15B-V1.0": ModelType.IFT,
+    "gpt2": ModelType.PT,
+    "keyfan/vicuna-chinese-replication-v1.1": ModelType.IFT,
+    "nthngdy/pythia-owt2-70m-100k": ModelType.FT,
+    "nthngdy/pythia-owt2-70m-50k": ModelType.FT,
+    "quantumaikr/KoreanLM-hf": ModelType.FT,
+    "quantumaikr/open_llama_7b_hf": ModelType.FT,
+    "quantumaikr/QuantumLM-70B-hf": ModelType.IFT,
+    "MayaPH/FinOPT-Lincoln": ModelType.FT,
+    "MayaPH/FinOPT-Franklin": ModelType.FT,
+    "MayaPH/GodziLLa-30B": ModelType.IFT,
+    "MayaPH/GodziLLa-30B-plus": ModelType.IFT,
+    "MayaPH/FinOPT-Washington": ModelType.FT,
+    "ogimgio/gpt-neo-125m-neurallinguisticpioneers": ModelType.FT,
+    "layoric/llama-2-13b-code-alpaca": ModelType.FT,
+    "CobraMamba/mamba-gpt-3b": ModelType.FT,
+    "CobraMamba/mamba-gpt-3b-v2": ModelType.FT,
+    "CobraMamba/mamba-gpt-3b-v3": ModelType.FT,
+    "timdettmers/guanaco-33b-merged": ModelType.FT,
+    "elinas/chronos-33b": ModelType.IFT,
+    "heegyu/RedTulu-Uncensored-3B-0719": ModelType.IFT,
+    "heegyu/WizardVicuna-Uncensored-3B-0719": ModelType.IFT,
+    "heegyu/WizardVicuna-3B-0719": ModelType.IFT,
+    "meta-llama/Llama-2-7b-chat-hf": ModelType.RL,
+    "meta-llama/Llama-2-7b-hf": ModelType.PT,
+    "meta-llama/Llama-2-13b-chat-hf": ModelType.RL,
+    "meta-llama/Llama-2-13b-hf": ModelType.PT,
+    "meta-llama/Llama-2-70b-chat-hf": ModelType.RL,
+    "meta-llama/Llama-2-70b-hf": ModelType.PT,
+    "xhyi/PT_GPTNEO350_ATG": ModelType.FT,
+    "h2oai/h2ogpt-gm-oasst1-en-1024-20b": ModelType.FT,
+    "h2oai/h2ogpt-gm-oasst1-en-1024-open-llama-7b-preview-400bt": ModelType.FT,
+    "h2oai/h2ogpt-oig-oasst1-512-6_9b": ModelType.IFT,
+    "h2oai/h2ogpt-oasst1-512-12b": ModelType.IFT,
+    "h2oai/h2ogpt-oig-oasst1-256-6_9b": ModelType.IFT,
+    "h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt": ModelType.FT,
+    "h2oai/h2ogpt-oasst1-512-20b": ModelType.IFT,
+    "h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b-preview-300bt-v2": ModelType.FT,
+    "h2oai/h2ogpt-gm-oasst1-en-1024-12b": ModelType.FT,
+    "h2oai/h2ogpt-gm-oasst1-multilang-1024-20b": ModelType.FT,
+    "bofenghuang/vigogne-13b-instruct": ModelType.IFT,
+    "bofenghuang/vigogne-13b-chat": ModelType.FT,
+    "bofenghuang/vigogne-2-7b-instruct": ModelType.IFT,
+    "bofenghuang/vigogne-7b-instruct": ModelType.IFT,
+    "bofenghuang/vigogne-7b-chat": ModelType.FT,
+    "Vmware/open-llama-7b-v2-open-instruct": ModelType.IFT,
+    "VMware/open-llama-0.7T-7B-open-instruct-v1.1": ModelType.IFT,
+    "ewof/koishi-instruct-3b": ModelType.IFT,
+    "gywy/llama2-13b-chinese-v1": ModelType.FT,
+    "GOAT-AI/GOAT-7B-Community": ModelType.FT,
+    "psyche/kollama2-7b": ModelType.FT,
+    "TheTravellingEngineer/llama2-7b-hf-guanaco": ModelType.FT,
+    "beaugogh/pythia-1.4b-deduped-sharegpt": ModelType.FT,
+    "augtoma/qCammel-70-x": ModelType.IFT,
+    "Lajonbot/Llama-2-7b-chat-hf-instruct-pl-lora_unload": ModelType.IFT,
+    "anhnv125/pygmalion-6b-roleplay": ModelType.FT,
+    "64bits/LexPodLM-13B": ModelType.FT,
+}
+
+
+def model_type_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

src/display_models/modelcard_filter.py

@@ -0,0 +1,26 @@
+import huggingface_hub
+from huggingface_hub import ModelCard
+
+
+# ht to @Wauplin, thank you for the snippet!
+# See https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/317
+def check_model_card(repo_id: str) -> tuple[bool, str]:
+    # Returns operation status, and error message
+    try:
+        card = ModelCard.load(repo_id)
+    except huggingface_hub.utils.EntryNotFoundError:
+        return False, "Please add a model card to your model to explain how you trained/fine-tuned it."
+
+    # 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, ""

src/display_models/read_results.py

@@ -0,0 +1,153 @@
+import json
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Tuple
+
+import dateutil
+import numpy as np
+
+from src.display_models.utils import AutoEvalColumn, make_clickable_model
+
+METRICS = ["acc_norm", "acc_norm", "acc", "mc2"]
+BENCHMARKS = ["arc:challenge", "hellaswag", "hendrycksTest", "truthfulqa:mc"]
+BENCH_TO_NAME = {
+    "arc:challenge": AutoEvalColumn.arc.name,
+    "hellaswag": AutoEvalColumn.hellaswag.name,
+    "hendrycksTest": AutoEvalColumn.mmlu.name,
+    "truthfulqa:mc": AutoEvalColumn.truthfulqa.name,
+}
+
+
+@dataclass
+class EvalResult:
+    eval_name: str
+    org: str
+    model: str
+    revision: str
+    results: dict
+    precision: str = ""
+    model_type: str = ""
+    weight_type: str = "Original"
+    date: str = ""
+
+    def to_dict(self):
+        from src.load_from_hub import is_model_on_hub
+
+        if self.org is not None:
+            base_model = f"{self.org}/{self.model}"
+        else:
+            base_model = f"{self.model}"
+        data_dict = {}
+
+        data_dict["eval_name"] = self.eval_name  # not a column, just a save name
+        data_dict["weight_type"] = self.weight_type  # not a column, just a save name
+        data_dict[AutoEvalColumn.precision.name] = self.precision
+        data_dict[AutoEvalColumn.model_type.name] = self.model_type
+        data_dict[AutoEvalColumn.model.name] = make_clickable_model(base_model)
+        data_dict[AutoEvalColumn.dummy.name] = base_model
+        data_dict[AutoEvalColumn.revision.name] = self.revision
+        data_dict[AutoEvalColumn.average.name] = sum([v for k, v in self.results.items()]) / 4.0
+        data_dict[AutoEvalColumn.still_on_hub.name] = (
+            is_model_on_hub(base_model, self.revision)[0] or base_model == "baseline"
+        )
+
+        for benchmark in BENCHMARKS:
+            if benchmark not in self.results.keys():
+                self.results[benchmark] = None
+
+        for k, v in BENCH_TO_NAME.items():
+            data_dict[v] = self.results[k]
+
+        return data_dict
+
+
+def parse_eval_result(json_filepath: str) -> Tuple[str, list[dict]]:
+    with open(json_filepath) as fp:
+        data = json.load(fp)
+
+    for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]:
+        if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0:
+            return None, []  # we skip models with the wrong version
+
+    try:
+        config = data["config"]
+    except KeyError:
+        config = data["config_general"]
+    model = config.get("model_name", None)
+    if model is None:
+        model = config.get("model_args", None)
+
+    model_sha = config.get("model_sha", "")
+    model_split = model.split("/", 1)
+
+    precision = config.get("model_dtype")
+
+    model = model_split[-1]
+
+    if len(model_split) == 1:
+        org = None
+        model = model_split[0]
+        result_key = f"{model}_{precision}"
+    else:
+        org = model_split[0]
+        model = model_split[1]
+        result_key = f"{org}_{model}_{precision}"
+
+    eval_results = []
+    for benchmark, metric in zip(BENCHMARKS, METRICS):
+        accs = np.array([v.get(metric, None) for k, v in data["results"].items() if benchmark in k])
+        if accs.size == 0 or any([acc is None for acc in accs]):
+            continue
+        mean_acc = np.mean(accs) * 100.0
+        eval_results.append(
+            EvalResult(
+                eval_name=result_key,
+                org=org,
+                model=model,
+                revision=model_sha,
+                results={benchmark: mean_acc},
+                precision=precision,  # todo model_type=, weight_type=
+                date=config.get("submission_date"),
+            )
+        )
+
+    return result_key, eval_results
+
+
+def get_eval_results() -> List[EvalResult]:
+    json_filepaths = []
+
+    for root, dir, files in os.walk("eval-results"):
+        # 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
+        # store results by precision maybe?
+        try:
+            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
+        except dateutil.parser._parser.ParserError:
+            files = [files[-1]]
+
+        # up_to_date = files[-1]
+        for file in files:
+            json_filepaths.append(os.path.join(root, file))
+
+    eval_results = {}
+    for json_filepath in json_filepaths:
+        result_key, results = parse_eval_result(json_filepath)
+        for eval_result in results:
+            if result_key in eval_results.keys():
+                eval_results[result_key].results.update(eval_result.results)
+            else:
+                eval_results[result_key] = eval_result
+
+    eval_results = [v for v in eval_results.values()]
+
+    return eval_results
+
+
+def get_eval_results_dicts() -> List[Dict]:
+    eval_results = get_eval_results()
+
+    return [e.to_dict() for e in eval_results]

src/display_models/utils.py

@@ -0,0 +1,146 @@
+import os
+from dataclasses import dataclass
+
+from huggingface_hub import HfApi
+
+API = HfApi()
+
+
+# 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
+
+
+def fields(raw_class):
+    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
+
+
+@dataclass(frozen=True)
+class AutoEvalColumn:  # Auto evals column
+    model_type_symbol = ColumnContent("T", "str", True)
+    model = ColumnContent("Model", "markdown", True)
+    average = ColumnContent("Average ⬆️", "number", True)
+    arc = ColumnContent("ARC", "number", True)
+    hellaswag = ColumnContent("HellaSwag", "number", True)
+    mmlu = ColumnContent("MMLU", "number", True)
+    truthfulqa = ColumnContent("TruthfulQA", "number", True)
+    model_type = ColumnContent("Type", "str", False)
+    precision = ColumnContent("Precision", "str", False)  # , True)
+    license = ColumnContent("Hub License", "str", False)
+    params = ColumnContent("#Params (B)", "number", False)
+    likes = ColumnContent("Hub ❤️", "number", False)
+    still_on_hub = ColumnContent("Available on the hub", "bool", False)
+    revision = ColumnContent("Model sha", "str", False, False)
+    dummy = ColumnContent(
+        "model_name_for_query", "str", True
+    )  # dummy col to implement search bar (hidden by custom CSS)
+
+
+@dataclass(frozen=True)
+class EloEvalColumn:  # Elo evals column
+    model = ColumnContent("Model", "markdown", True)
+    gpt4 = ColumnContent("GPT-4 (all)", "number", True)
+    human_all = ColumnContent("Human (all)", "number", True)
+    human_instruct = ColumnContent("Human (instruct)", "number", True)
+    human_code_instruct = ColumnContent("Human (code-instruct)", "number", True)
+
+
+@dataclass(frozen=True)
+class EvalQueueColumn:  # Queue column
+    model = ColumnContent("model", "markdown", True)
+    revision = ColumnContent("revision", "str", True)
+    private = ColumnContent("private", "bool", True)
+    precision = ColumnContent("precision", "str", True)
+    weight_type = ColumnContent("weight_type", "str", "Original")
+    status = ColumnContent("status", "str", True)
+
+
+LLAMAS = [
+    "huggingface/llama-7b",
+    "huggingface/llama-13b",
+    "huggingface/llama-30b",
+    "huggingface/llama-65b",
+]
+
+
+KOALA_LINK = "https://huggingface.co/TheBloke/koala-13B-HF"
+VICUNA_LINK = "https://huggingface.co/lmsys/vicuna-13b-delta-v1.1"
+OASST_LINK = "https://huggingface.co/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5"
+DOLLY_LINK = "https://huggingface.co/databricks/dolly-v2-12b"
+MODEL_PAGE = "https://huggingface.co/models"
+LLAMA_LINK = "https://ai.facebook.com/blog/large-language-model-llama-meta-ai/"
+VICUNA_LINK = "https://huggingface.co/CarperAI/stable-vicuna-13b-delta"
+ALPACA_LINK = "https://crfm.stanford.edu/2023/03/13/alpaca.html"
+
+
+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}"
+
+    if model_name in LLAMAS:
+        link = LLAMA_LINK
+        model_name = model_name.split("/")[1]
+    elif model_name == "HuggingFaceH4/stable-vicuna-13b-2904":
+        link = VICUNA_LINK
+        model_name = "stable-vicuna-13b"
+    elif model_name == "HuggingFaceH4/llama-7b-ift-alpaca":
+        link = ALPACA_LINK
+        model_name = "alpaca-13b"
+    if model_name == "dolly-12b":
+        link = DOLLY_LINK
+    elif model_name == "vicuna-13b":
+        link = VICUNA_LINK
+    elif model_name == "koala-13b":
+        link = KOALA_LINK
+    elif model_name == "oasst-12b":
+        link = OASST_LINK
+
+    details_model_name = model_name.replace("/", "__")
+    details_link = f"https://huggingface.co/datasets/open-llm-leaderboard/details_{details_model_name}"
+
+    if not bool(os.getenv("DEBUG", "False")):
+        # We only add these checks when not debugging, as they are extremely slow
+        print(f"details_link: {details_link}")
+        try:
+            check_path = list(
+                API.list_files_info(
+                    repo_id=f"open-llm-leaderboard/details_{details_model_name}",
+                    paths="README.md",
+                    repo_type="dataset",
+                )
+            )
+            print(f"check_path: {check_path}")
+        except Exception as err:
+            # No details repo for this model
+            print(f"No details repo for this model: {err}")
+            return model_hyperlink(link, model_name)
+
+    return model_hyperlink(link, model_name) + "  " + model_hyperlink(details_link, "📑")
+
+
+def styled_error(error):
+    return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"
+
+
+def styled_warning(warn):
+    return f"<p style='color: orange; font-size: 20px; text-align: center;'>{warn}</p>"
+
+
+def styled_message(message):
+    return f"<p style='color: green; font-size: 20px; text-align: center;'>{message}</p>"
+
+
+def has_no_nan_values(df, columns):
+    return df[columns].notna().all(axis=1)
+
+
+def has_nan_values(df, columns):
+    return df[columns].isna().any(axis=1)

src/envs.py

@@ -1,46 +0,0 @@
-import os
-import logging
-
-from huggingface_hub import HfApi
-
-# clone / pull the lmeh eval data
-H4_TOKEN = os.environ.get("H4_TOKEN", None)
-
-REPO_ID = "HuggingFaceH4/open_llm_leaderboard"
-QUEUE_REPO = "open-llm-leaderboard/requests"
-DYNAMIC_INFO_REPO = "open-llm-leaderboard/dynamic_model_information"
-RESULTS_REPO = "open-llm-leaderboard/results"
-
-PRIVATE_QUEUE_REPO = "open-llm-leaderboard/private-requests"
-PRIVATE_RESULTS_REPO = "open-llm-leaderboard/private-results"
-
-IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
-
-HF_HOME = os.getenv("HF_HOME", ".")
-
-# Check HF_HOME write access
-print(f"Initial HF_HOME set to: {HF_HOME}")
-
-if not os.access(HF_HOME, os.W_OK):
-    print(f"No write access to HF_HOME: {HF_HOME}. Resetting to current directory.")
-    HF_HOME = "."
-    os.environ["HF_HOME"] = HF_HOME
-else:
-    print("Write access confirmed for HF_HOME")
-
-EVAL_REQUESTS_PATH = os.path.join(HF_HOME, "eval-queue")
-EVAL_RESULTS_PATH = os.path.join(HF_HOME, "eval-results")
-DYNAMIC_INFO_PATH = os.path.join(HF_HOME, "dynamic-info")
-DYNAMIC_INFO_FILE_PATH = os.path.join(DYNAMIC_INFO_PATH, "model_infos.json")
-
-EVAL_REQUESTS_PATH_PRIVATE = "eval-queue-private"
-EVAL_RESULTS_PATH_PRIVATE = "eval-results-private"
-
-PATH_TO_COLLECTION = "open-llm-leaderboard/llm-leaderboard-best-models-652d6c7965a4619fb5c27a03"
-
-# Rate limit variables
-RATE_LIMIT_PERIOD = 7
-RATE_LIMIT_QUOTA = 5
-HAS_HIGHER_RATE_LIMIT = ["TheBloke"]
-
-API = HfApi(token=H4_TOKEN)

src/leaderboard/filter_models.py

@@ -1,171 +0,0 @@
-from src.display.formatting import model_hyperlink
-from src.display.utils import AutoEvalColumn
-
-
-# Models which have been flagged by users as being problematic for a reason or another
-# (Model name to forum discussion link)
-FLAGGED_MODELS = {
-    "merged": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Voicelab/trurl-2-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/202",
-    "deepnight-research/llama-2-70B-inst": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/207",
-    "Aspik101/trurl-2-13b-pl-instruct_unload": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/213",
-    "Fredithefish/ReasonixPajama-3B-HF": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/236",
-    "TigerResearch/tigerbot-7b-sft-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/237",
-    "gaodrew/gaodrew-gorgonzola-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/215",
-    "AIDC-ai-business/Marcoroni-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
-    "AIDC-ai-business/Marcoroni-13B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
-    "AIDC-ai-business/Marcoroni-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/287",
-    "fblgit/una-xaberius-34b-v1beta": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/444",
-    "jan-hq/trinity-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "rwitz2/go-bruins-v2.1.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "rwitz2/go-bruins-v2.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "GreenNode/GreenNodeLM-v3olet-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "GreenNode/GreenNodeLM-7B-v4leo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "GreenNode/LeoScorpius-GreenNode-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "viethq188/LeoScorpius-7B-Chat-DPO": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "GreenNode/GreenNodeLM-7B-v2leo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "janai-hq/trinity-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "ignos/LeoScorpius-GreenNode-Alpaca-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "fblgit/una-cybertron-7b-v3-OMA": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "mncai/mistral-7b-dpo-merge-v1.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "mncai/mistral-7b-dpo-v6": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "Toten5/LeoScorpius-GreenNode-7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "GreenNode/GreenNodeLM-7B-v1olet": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "quantumaikr/quantum-dpo-v0.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "quantumaikr/quantum-v0.01": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "quantumaikr/quantum-trinity-v0.1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "mncai/mistral-7b-dpo-v5": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "cookinai/BruinHermes": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "jan-ai/Pandora-10.7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "v1olet/v1olet_marcoroni-go-bruins-merge-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "v1olet/v1olet_merged_dpo_7B_v3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "rwitz2/pee": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "zyh3826 / GML-Mistral-merged-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/503",
-    "dillfrescott/trinity-medium": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/474",
-    "udkai/Garrulus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/526",
-    "dfurman/GarrulusMarcoro-7B-v0.1": "https://huggingface.co/dfurman/GarrulusMarcoro-7B-v0.1/discussions/1",
-    "eren23/slerp-test-turdus-beagle": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    "abideen/NexoNimbus-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    "alnrg2arg/test2_3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    "nfaheem/Marcoroni-7b-DPO-Merge": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    "CultriX/MergeTrix-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    "liminerity/Blur-7b-v1.21": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/548",
-    # Merges not indicated
-    "gagan3012/MetaModelv2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "gagan3012/MetaModelv3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "kyujinpy/Sakura-SOLRCA-Math-Instruct-DPO-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "kyujinpy/Sakura-SOLAR-Instruct-DPO-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "kyujinpy/Sakura-SOLRCA-Math-Instruct-DPO-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "kyujinpy/Sakura-SOLRCA-Instruct-DPO": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "fblgit/LUNA-SOLARkrautLM-Instruct": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "perlthoughts/Marcoroni-8x7B-v3-MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "rwitz/go-bruins-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "rwitz/go-bruins": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Walmart-the-bag/Solar-10.7B-Cato": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "aqweteddy/mistral_tv-neural-marconroni": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "NExtNewChattingAI/shark_tank_ai_7_b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Q-bert/MetaMath-Cybertron": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "OpenPipe/mistral-ft-optimized-1227": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "perlthoughts/Falkor-7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "v1olet/v1olet_merged_dpo_7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Ba2han/BruinsV2-OpHermesNeu-11B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "DopeorNope/You_can_cry_Snowman-13B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "PistachioAlt/Synatra-MCS-7B-v0.3-RP-Slerp": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/MetaMath-una-cybertron-v2-bf16-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/OpenHermes-2.5-neural-chat-7b-v3-2-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "perlthoughts/Falkor-8x7B-MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "elinas/chronos007-70b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/MetaMath-NeuralHermes-2.5-Mistral-7B-Linear": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/MetaMath-neural-chat-7b-v3-2-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "diffnamehard/Mistral-CatMacaroni-slerp-uncensored-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/neural-chat-7b-v3-1-OpenHermes-2.5-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/MetaMath-NeuralHermes-2.5-Mistral-7B-Ties": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Walmart-the-bag/Misted-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "garage-bAInd/Camel-Platypus2-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "Weyaxi/OpenOrca-Zephyr-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "uukuguy/speechless-mistral-7b-dare-0.85": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/510",
-    "DopeorNope/SOLARC-M-10.7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
-    "cloudyu/Mixtral_11Bx2_MoE_19B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
-    "DopeorNope/SOLARC-MOE-10.7Bx6 ": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
-    "DopeorNope/SOLARC-MOE-10.7Bx4": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
-    "gagan3012/MetaModelv2 ": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/511",
-    "udkai/Turdus": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "kodonho/Solar-OrcaDPO-Solar-Instruct-SLERP": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "kodonho/SolarM-SakuraSolar-SLERP": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "Yhyu13/LMCocktail-10.7B-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "mlabonne/NeuralMarcoro14-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "Neuronovo/neuronovo-7B-v0.2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "ryandt/MusingCaterpillar": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "Neuronovo/neuronovo-7B-v0.3": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "SanjiWatsuki/Lelantos-DPO-7B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "bardsai/jaskier-7b-dpo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "cookinai/OpenCM-14": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "bardsai/jaskier-7b-dpo-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "jan-hq/supermario-v2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    # MoErges
-    "cloudyu/Yi-34Bx2-MoE-60B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "cloudyu/Mixtral_34Bx2_MoE_60B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "gagan3012/MetaModel_moe": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/SOLAR-math-2x10.7b-v0.2": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "cloudyu/Mixtral_7Bx2_MoE": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/SOLAR-math-2x10.7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/Orca-SOLAR-4x10.7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/piccolo-8x7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "cloudyu/Mixtral_7Bx4_MOE_24B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/laser-dolphin-mixtral-2x7b-dpo": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    "macadeliccc/polyglot-math-4x7b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/540",
-    # Other - contamination mostly
-    "DopeorNope/COKAL-v1-70B": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/566",
-    "CultriX/MistralTrix-v1": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/556",
-    "Contamination/contaminated_proof_7b_v1.0": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/664",
-    "Contamination/contaminated_proof_7b_v1.0_safetensor": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/664",
-}
-
-# Models which have been requested by orgs to not be submitted on the leaderboard
-DO_NOT_SUBMIT_MODELS = [
-    "Voicelab/trurl-2-13b",  # trained on MMLU
-    "TigerResearch/tigerbot-70b-chat",  # per authors request
-    "TigerResearch/tigerbot-70b-chat-v2",  # per authors request
-    "TigerResearch/tigerbot-70b-chat-v4-4k",  # per authors request
-]
-
-
-def flag_models(leaderboard_data: list[dict]):
-    """Flags models based on external criteria or flagged status."""
-    for model_data in leaderboard_data:
-        # Merges and moes are flagged automatically
-        if model_data[AutoEvalColumn.flagged.name]:
-            flag_key = "merged"
-        else:
-            flag_key = model_data[AutoEvalColumn.fullname.name]
-        if flag_key in FLAGGED_MODELS:
-            issue_num = FLAGGED_MODELS[flag_key].split("/")[-1]
-            issue_link = model_hyperlink(
-                FLAGGED_MODELS[flag_key],
-                f"See discussion #{issue_num}",
-            )
-            model_data[AutoEvalColumn.model.name] = (
-                f"{model_data[AutoEvalColumn.model.name]} has been flagged! {issue_link}"
-            )
-            model_data[AutoEvalColumn.flagged.name] = True
-        else:
-            model_data[AutoEvalColumn.flagged.name] = False
-
-
-def remove_forbidden_models(leaderboard_data: list[dict]):
-    """Removes models from the leaderboard based on the DO_NOT_SUBMIT list."""
-    indices_to_remove = []
-    for ix, model in enumerate(leaderboard_data):
-        if model[AutoEvalColumn.fullname.name] in DO_NOT_SUBMIT_MODELS:
-            indices_to_remove.append(ix)
-
-    # Remove the models from the list
-    for ix in reversed(indices_to_remove):
-        leaderboard_data.pop(ix)
-    return leaderboard_data
-
-
-def filter_models_flags(leaderboard_data: list[dict]):
-    leaderboard_data = remove_forbidden_models(leaderboard_data)
-    flag_models(leaderboard_data)
-

src/leaderboard/read_evals.py

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

src/load_from_hub.py

@@ -0,0 +1,152 @@
+import json
+import os
+
+import pandas as pd
+from huggingface_hub import Repository
+from transformers import AutoConfig
+from collections import defaultdict
+
+from src.assets.hardcoded_evals import baseline, gpt4_values, gpt35_values
+from src.display_models.get_model_metadata import apply_metadata
+from src.display_models.read_results import get_eval_results_dicts, make_clickable_model
+from src.display_models.utils import AutoEvalColumn, EvalQueueColumn, has_no_nan_values
+
+IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
+
+
+def get_all_requested_models(requested_models_dir: str) -> set[str]:
+    depth = 1
+    file_names = []
+    users_to_submission_dates = defaultdict(list)
+
+    for root, _, files in os.walk(requested_models_dir):
+        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
+        if current_depth == depth:
+            for file in files:
+                if not file.endswith(".json"):
+                    continue
+                with open(os.path.join(root, file), "r") as f:
+                    info = json.load(f)
+                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
+
+                    # Select organisation
+                    if info["model"].count("/") == 0 or "submitted_time" not in info:
+                        continue
+                    organisation, _ = info["model"].split("/")
+                    users_to_submission_dates[organisation].append(info["submitted_time"])
+
+    return set(file_names), users_to_submission_dates
+
+
+def load_all_info_from_hub(QUEUE_REPO: str, RESULTS_REPO: str, QUEUE_PATH: str, RESULTS_PATH: str) -> list[Repository]:
+    eval_queue_repo = None
+    eval_results_repo = None
+    requested_models = None
+
+    print("Pulling evaluation requests and results.")
+
+    eval_queue_repo = Repository(
+        local_dir=QUEUE_PATH,
+        clone_from=QUEUE_REPO,
+        repo_type="dataset",
+    )
+    eval_queue_repo.git_pull()
+
+    eval_results_repo = Repository(
+        local_dir=RESULTS_PATH,
+        clone_from=RESULTS_REPO,
+        repo_type="dataset",
+    )
+    eval_results_repo.git_pull()
+
+    requested_models, users_to_submission_dates = get_all_requested_models("eval-queue")
+
+    return eval_queue_repo, requested_models, eval_results_repo, users_to_submission_dates
+
+
+def get_leaderboard_df(
+    eval_results: Repository, eval_results_private: Repository, cols: list, benchmark_cols: list
+) -> pd.DataFrame:
+    if eval_results:
+        print("Pulling evaluation results for the leaderboard.")
+        eval_results.git_pull()
+    if eval_results_private:
+        print("Pulling evaluation results for the leaderboard.")
+        eval_results_private.git_pull()
+
+    all_data = get_eval_results_dicts()
+
+    if not IS_PUBLIC:
+        all_data.append(gpt4_values)
+        all_data.append(gpt35_values)
+
+    all_data.append(baseline)
+    apply_metadata(all_data)  # Populate model type based on known hardcoded values in `metadata.py`
+
+    df = pd.DataFrame.from_records(all_data)
+    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(
+    eval_queue: Repository, eval_queue_private: Repository, save_path: str, cols: list
+) -> list[pd.DataFrame]:
+    if eval_queue:
+        print("Pulling changes for the evaluation queue.")
+        eval_queue.git_pull()
+    if eval_queue_private:
+        print("Pulling changes for the evaluation queue.")
+        eval_queue_private.git_pull()
+
+    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 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 is_model_on_hub(model_name: str, revision: str) -> bool:
+    try:
+        AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=False)
+        return True, None
+
+    except ValueError:
+        return (
+            False,
+            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
+        )
+
+    except Exception as e:
+        print(f"Could not get the model config from the hub.: {e}")
+        return False, "was not found on hub!"

src/populate.py

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

src/rate_limiting.py

@@ -0,0 +1,13 @@
+from datetime import datetime, timezone, timedelta
+
+
+def user_submission_permission(submission_name, users_to_submission_dates, rate_limit_period):
+    org_or_user, _ = submission_name.split("/")
+    if org_or_user not in users_to_submission_dates:
+        return 0
+    submission_dates = sorted(users_to_submission_dates[org_or_user])
+
+    time_limit = (datetime.now(timezone.utc) - timedelta(days=rate_limit_period)).strftime("%Y-%m-%dT%H:%M:%SZ")
+    submissions_after_timelimit = [d for d in submission_dates if d > time_limit]
+
+    return len(submissions_after_timelimit)

src/scripts/create_request_file.py

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

src/scripts/update_all_request_files.py

@@ -1,129 +0,0 @@
-import json
-import os
-import time
-
-from huggingface_hub import snapshot_download
-
-from src.envs import API, DYNAMIC_INFO_FILE_PATH, DYNAMIC_INFO_PATH, DYNAMIC_INFO_REPO, EVAL_REQUESTS_PATH, H4_TOKEN
-from src.submission.check_validity import check_model_card, get_model_tags, is_model_on_hub
-
-
-def update_one_model(model_id, data, models_on_the_hub):
-    # Model no longer on the hub at all
-    if model_id not in models_on_the_hub:
-        data["still_on_hub"] = False
-        data["likes"] = 0
-        data["downloads"] = 0
-        data["created_at"] = ""
-        data["tags"] = []
-        return data
-
-    # Grabbing model parameters
-    model_cfg = models_on_the_hub[model_id]
-    data["likes"] = model_cfg.likes
-    data["downloads"] = model_cfg.downloads
-    data["created_at"] = str(model_cfg.created_at)
-    data["license"] = model_cfg.card_data.license if model_cfg.card_data is not None else ""
-
-    # Grabbing model details
-    model_name = model_id
-    if model_cfg.card_data is not None and model_cfg.card_data.base_model is not None:
-        if isinstance(model_cfg.card_data.base_model, str):
-            model_name = model_cfg.card_data.base_model  # for adapters, we look at the parent model
-    still_on_hub, _, _ = is_model_on_hub(
-        model_name=model_name,
-        revision=data.get("revision"),
-        trust_remote_code=True,
-        test_tokenizer=False,
-        token=H4_TOKEN,
-    )
-    # If the model doesn't have a model card or a license, we consider it's deleted
-    if still_on_hub:
-        try:
-            status, _, model_card = check_model_card(model_id)
-            if status is False:
-                still_on_hub = False
-        except Exception:
-            model_card = None
-            still_on_hub = False
-    data["still_on_hub"] = still_on_hub
-
-    tags = get_model_tags(model_card, model_id) if still_on_hub else []
-
-    data["tags"] = tags
-    return data
-
-
-def update_models(file_path, models_on_the_hub):
-    """
-    Search through all JSON files in the specified root folder and its subfolders,
-    and update the likes key in JSON dict from value of input dict
-    """
-    seen_models = []
-    with open(file_path, "r") as f:
-        model_infos = json.load(f)
-        for model_id in model_infos.keys():
-            seen_models.append(model_id)
-            model_infos[model_id] = update_one_model(
-                model_id=model_id, data=model_infos[model_id], models_on_the_hub=models_on_the_hub
-            )
-
-    # If new requests files have been created since we started all this
-    # we grab them
-    all_models = []
-    try:
-        for ix, (root, _, files) in enumerate(os.walk(EVAL_REQUESTS_PATH)):
-            if ix == 0:
-                continue
-            for file in files:
-                if "eval_request" in file:
-                    path = root.split("/")[-1] + "/" + file.split("_eval_request")[0]
-                    all_models.append(path)
-    except Exception as e:
-        print(e)
-        pass
-
-    for model_id in all_models:
-        if model_id not in seen_models:
-            model_infos[model_id] = update_one_model(model_id=model_id, data={}, models_on_the_hub=models_on_the_hub)
-
-    with open(file_path, "w") as f:
-        json.dump(model_infos, f, indent=2)
-
-
-def update_dynamic_files():
-    """This will only update metadata for models already linked in the repo, not add missing ones."""
-    snapshot_download(
-        repo_id=DYNAMIC_INFO_REPO, local_dir=DYNAMIC_INFO_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
-    )
-
-    print("UPDATE_DYNAMIC: Loaded snapshot")
-    # Get models
-    start = time.time()
-
-    models = list(
-        API.list_models(
-            # filter=ModelFilter(task="text-generation"),
-            full=False,
-            cardData=True,
-            fetch_config=True,
-        )
-    )
-    id_to_model = {model.id: model for model in models}
-
-    print(f"UPDATE_DYNAMIC: Downloaded list of models in {time.time() - start:.2f} seconds")
-
-    start = time.time()
-
-    update_models(DYNAMIC_INFO_FILE_PATH, id_to_model)
-
-    print(f"UPDATE_DYNAMIC: updated in {time.time() - start:.2f} seconds")
-
-    API.upload_file(
-        path_or_fileobj=DYNAMIC_INFO_FILE_PATH,
-        path_in_repo=DYNAMIC_INFO_FILE_PATH.split("/")[-1],
-        repo_id=DYNAMIC_INFO_REPO,
-        repo_type="dataset",
-        commit_message="Daily request file update.",
-    )
-    print("UPDATE_DYNAMIC: pushed to hub")

src/submission/check_validity.py

@@ -1,179 +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, get_safetensors_metadata
-from transformers import AutoConfig, AutoTokenizer
-
-from src.envs import HAS_HIGHER_RATE_LIMIT
-
-
-# ht to @Wauplin, thank you for the snippet!
-# See https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/317
-def check_model_card(repo_id: str) -> tuple[bool, str]:
-    # Returns operation status, and error message
-    try:
-        card = ModelCard.load(repo_id)
-    except huggingface_hub.utils.EntryNotFoundError:
-        return False, "Please add a model card to your model to explain how you trained/fine-tuned it.", None
-
-    # Enforce license metadata
-    if card.data.license is None:
-        if not ("license_name" in card.data and "license_link" in card.data):
-            return (
-                False,
-                (
-                    "License not found. Please add a license to your model card using the `license` metadata or a"
-                    " `license_name`/`license_link` pair."
-                ),
-                None,
-            )
-
-    # Enforce card content
-    if len(card.text) < 200:
-        return False, "Please add a description to your model card, it is too short.", None
-
-    return True, "", card
-
-
-def is_model_on_hub(
-    model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False
-) -> tuple[bool, str, AutoConfig]:
-    try:
-        config = AutoConfig.from_pretrained(
-            model_name, revision=revision, trust_remote_code=trust_remote_code, token=token
-        )  # , force_download=True)
-        if test_tokenizer:
-            try:
-                tk = AutoTokenizer.from_pretrained(
-                    model_name, revision=revision, trust_remote_code=trust_remote_code, token=token
-                )
-            except ValueError as e:
-                return (False, f"uses a tokenizer which is not in a transformers release: {e}", None)
-            except Exception:
-                return (
-                    False,
-                    "'s tokenizer cannot be loaded. Is your tokenizer class in a stable transformers release, and correctly configured?",
-                    None,
-                )
-        return True, None, config
-
-    except ValueError:
-        return (
-            False,
-            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
-            None,
-        )
-
-    except Exception as e:
-        if "You are trying to access a gated repo." in str(e):
-            return True, "uses a gated model.", None
-        return False, f"was not found or misconfigured on the hub! Error raised was {e.args[0]}", None
-
-
-def get_model_size(model_info: ModelInfo, precision: str):
-    size_pattern = re.compile(r"(\d+\.)?\d+(b|m)")
-    safetensors = None
-    try:
-        safetensors = get_safetensors_metadata(model_info.id)
-    except Exception as e:
-        print(e)
-
-    if safetensors is not None:
-        model_size = round(sum(safetensors.parameter_count.values()) / 1e9, 3)
-    else:
-        try:
-            size_match = re.search(size_pattern, model_info.id.lower())
-            model_size = size_match.group(0)
-            model_size = round(float(model_size[:-1]) if model_size[-1] == "b" else float(model_size[:-1]) / 1e3, 3)
-        except AttributeError:
-            return 0  # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py
-
-    size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.id.lower()) else 1
-    model_size = size_factor * model_size
-    return model_size
-
-
-def get_model_arch(model_info: ModelInfo):
-    return model_info.config.get("architectures", "Unknown")
-
-
-def user_submission_permission(org_or_user, users_to_submission_dates, rate_limit_period, rate_limit_quota):
-    if org_or_user not in users_to_submission_dates:
-        return True, ""
-    submission_dates = sorted(users_to_submission_dates[org_or_user])
-
-    time_limit = (datetime.now(timezone.utc) - timedelta(days=rate_limit_period)).strftime("%Y-%m-%dT%H:%M:%SZ")
-    submissions_after_timelimit = [d for d in submission_dates if d > time_limit]
-
-    num_models_submitted_in_period = len(submissions_after_timelimit)
-    if org_or_user in HAS_HIGHER_RATE_LIMIT:
-        rate_limit_quota = 2 * rate_limit_quota
-
-    if num_models_submitted_in_period > rate_limit_quota:
-        error_msg = f"Organisation or user `{org_or_user}`"
-        error_msg += f"already has {num_models_submitted_in_period} model requests submitted to the leaderboard "
-        error_msg += f"in the last {rate_limit_period} days.\n"
-        error_msg += (
-            "Please wait a couple of days before resubmitting, so that everybody can enjoy using the leaderboard 🤗"
-        )
-        return False, error_msg
-    return True, ""
-
-
-def already_submitted_models(requested_models_dir: str) -> set[str]:
-    depth = 1
-    file_names = []
-    users_to_submission_dates = defaultdict(list)
-
-    for root, _, files in os.walk(requested_models_dir):
-        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
-        if current_depth == depth:
-            for file in files:
-                if not file.endswith(".json"):
-                    continue
-                with open(os.path.join(root, file), "r") as f:
-                    info = json.load(f)
-                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
-
-                    # Select organisation
-                    if info["model"].count("/") == 0 or "submitted_time" not in info:
-                        continue
-                    organisation, _ = info["model"].split("/")
-                    users_to_submission_dates[organisation].append(info["submitted_time"])
-
-    return set(file_names), users_to_submission_dates
-
-
-def get_model_tags(model_card, model: str):
-    is_merge_from_metadata = False
-    is_moe_from_metadata = False
-
-    tags = []
-    if model_card is None:
-        return tags
-    if model_card.data.tags:
-        is_merge_from_metadata = any(
-            [tag in model_card.data.tags for tag in ["merge", "moerge", "mergekit", "lazymergekit"]]
-        )
-        is_moe_from_metadata = any([tag in model_card.data.tags for tag in ["moe", "moerge"]])
-
-    is_merge_from_model_card = any(
-        keyword in model_card.text.lower() for keyword in ["merged model", "merge model", "moerge"]
-    )
-    if is_merge_from_model_card or is_merge_from_metadata:
-        tags.append("merge")
-    is_moe_from_model_card = any(keyword in model_card.text.lower() for keyword in ["moe", "mixtral"])
-    # Hardcoding because of gating problem
-    if "Qwen/Qwen1.5-32B" in model:
-        print("HERE NSHJNKJSNJLAS")
-        is_moe_from_model_card = False
-    is_moe_from_name = "moe" in model.lower().replace("/", "-").replace("_", "-").split("-")
-    if is_moe_from_model_card or is_moe_from_name or is_moe_from_metadata:
-        tags.append("moe")
-
-    return tags

src/submission/submit.py

@@ -1,196 +0,0 @@
-import json
-import os
-from datetime import datetime, timezone
-
-from huggingface_hub import snapshot_download
-
-from src.display.formatting import styled_error, styled_message, styled_warning
-from src.envs import (
-    API,
-    DYNAMIC_INFO_FILE_PATH,
-    DYNAMIC_INFO_PATH,
-    DYNAMIC_INFO_REPO,
-    EVAL_REQUESTS_PATH,
-    H4_TOKEN,
-    QUEUE_REPO,
-    RATE_LIMIT_PERIOD,
-    RATE_LIMIT_QUOTA,
-)
-from src.leaderboard.filter_models import DO_NOT_SUBMIT_MODELS
-from src.submission.check_validity import (
-    already_submitted_models,
-    check_model_card,
-    get_model_size,
-    get_model_tags,
-    is_model_on_hub,
-    user_submission_permission,
-)
-
-REQUESTED_MODELS = None
-USERS_TO_SUBMISSION_DATES = None
-
-
-def add_new_eval(
-    model: str,
-    base_model: str,
-    revision: str,
-    precision: str,
-    private: bool,
-    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.")
-
-    # Is the user rate limited?
-    if user_name != "":
-        user_can_submit, error_msg = user_submission_permission(
-            user_name, USERS_TO_SUBMISSION_DATES, RATE_LIMIT_PERIOD, RATE_LIMIT_QUOTA
-        )
-        if not user_can_submit:
-            return styled_error(error_msg)
-
-    # Did the model authors forbid its submission to the leaderboard?
-    if model in DO_NOT_SUBMIT_MODELS or base_model in DO_NOT_SUBMIT_MODELS:
-        return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.")
-
-    if model == "CohereForAI/c4ai-command-r-plus":
-        return styled_warning(
-            "This model cannot be submitted manually on the leaderboard before the transformers release."
-        )
-
-    # Does the model actually exist?
-    if revision == "":
-        revision = "main"
-
-    # Is the model on the hub?
-    if weight_type in ["Delta", "Adapter"]:
-        base_model_on_hub, error, _ = is_model_on_hub(
-            model_name=base_model, revision=revision, token=H4_TOKEN, test_tokenizer=True
-        )
-        if not base_model_on_hub:
-            return styled_error(f'Base model "{base_model}" {error}')
-
-    architecture = "?"
-    downloads = 0
-    created_at = ""
-    if not weight_type == "Adapter":
-        model_on_hub, error, model_config = is_model_on_hub(model_name=model, revision=revision, test_tokenizer=True)
-        if not model_on_hub or model_config is None:
-            return styled_error(f'Model "{model}" {error}')
-        if model_config is not None:
-            architectures = getattr(model_config, "architectures", None)
-            if architectures:
-                architecture = ";".join(architectures)
-            downloads = getattr(model_config, "downloads", 0)
-            created_at = getattr(model_config, "created_at", "")
-
-    # Is the model info correctly filled?
-    try:
-        model_info = API.model_info(repo_id=model, revision=revision)
-    except Exception:
-        return styled_error("Could not get your model information. Please fill it up properly.")
-
-    model_size = get_model_size(model_info=model_info, precision=precision)
-
-    # Were the model card and license filled?
-    try:
-        license = model_info.cardData["license"]
-    except Exception:
-        return styled_error("Please select a license for your model")
-
-    modelcard_OK, error_msg, model_card = check_model_card(model)
-    if not modelcard_OK:
-        return styled_error(error_msg)
-
-    tags = get_model_tags(model_card, model)
-
-    # Seems good, creating the eval
-    print("Adding new eval")
-
-    eval_entry = {
-        "model": model,
-        "base_model": base_model,
-        "revision": revision,
-        "private": private,
-        "precision": precision,
-        "params": model_size,
-        "architectures": architecture,
-        "weight_type": weight_type,
-        "status": "PENDING",
-        "submitted_time": current_time,
-        "model_type": model_type,
-        "job_id": -1,
-        "job_start_time": None,
-    }
-
-    supplementary_info = {
-        "likes": model_info.likes,
-        "license": license,
-        "still_on_hub": True,
-        "tags": tags,
-        "downloads": downloads,
-        "created_at": created_at,
-    }
-
-    # Check for duplicate submission
-    if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
-        return styled_warning("This model has been already submitted.")
-
-    print("Creating eval file")
-    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
-    os.makedirs(OUT_DIR, exist_ok=True)
-    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json"
-
-    with open(out_path, "w") as f:
-        f.write(json.dumps(eval_entry))
-
-    print("Uploading eval file")
-    API.upload_file(
-        path_or_fileobj=out_path,
-        path_in_repo=out_path.split("eval-queue/")[1],
-        repo_id=QUEUE_REPO,
-        repo_type="dataset",
-        commit_message=f"Add {model} to eval queue",
-    )
-
-    # We want to grab the latest version of the submission file to not accidentally overwrite it
-    snapshot_download(
-        repo_id=DYNAMIC_INFO_REPO, local_dir=DYNAMIC_INFO_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
-    )
-
-    with open(DYNAMIC_INFO_FILE_PATH) as f:
-        all_supplementary_info = json.load(f)
-
-    all_supplementary_info[model] = supplementary_info
-    with open(DYNAMIC_INFO_FILE_PATH, "w") as f:
-        json.dump(all_supplementary_info, f, indent=2)
-
-    API.upload_file(
-        path_or_fileobj=DYNAMIC_INFO_FILE_PATH,
-        path_in_repo=DYNAMIC_INFO_FILE_PATH.split("/")[-1],
-        repo_id=DYNAMIC_INFO_REPO,
-        repo_type="dataset",
-        commit_message=f"Add {model} to dynamic info queue",
-    )
-
-    # Remove the local file
-    os.remove(out_path)
-
-    return styled_message(
-        "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
-    )

src/tools/collections.py

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

src/tools/plots.py

@@ -1,159 +0,0 @@
-import numpy as np
-import pandas as pd
-import plotly.express as px
-from plotly.graph_objs import Figure
-
-from src.display.utils import BENCHMARK_COLS, AutoEvalColumn, Task, Tasks
-from src.display.utils import human_baseline_row as HUMAN_BASELINE
-from src.leaderboard.filter_models import FLAGGED_MODELS
-from src.leaderboard.read_evals import EvalResult
-
-
-def create_scores_df(raw_data: list[EvalResult]) -> pd.DataFrame:
-    """
-    Generates a DataFrame containing the maximum scores until each date.
-
-    :param results_df: A DataFrame containing result information including metric scores and dates.
-    :return: A new DataFrame containing the maximum scores until each date for every metric.
-    """
-    # Step 1: Ensure 'date' is in datetime format and sort the DataFrame by it
-    results_df = pd.DataFrame(raw_data)
-    # results_df["date"] = pd.to_datetime(results_df["date"], format="mixed", utc=True)
-    results_df.sort_values(by="date", inplace=True)
-
-    # Step 2: Initialize the scores dictionary
-    scores = {k: [] for k in BENCHMARK_COLS + [AutoEvalColumn.average.name]}
-
-    # Step 3: Iterate over the rows of the DataFrame and update the scores dictionary
-    for task in [t.value for t in Tasks] + [Task("Average", "avg", AutoEvalColumn.average.name)]:
-        current_max = 0
-        last_date = ""
-        column = task.col_name
-        for _, row in results_df.iterrows():
-            current_model = row["full_model"]
-            # We ignore models that are flagged/no longer on the hub/not finished
-            to_ignore = (
-                not row["still_on_hub"]
-                or row["flagged"]
-                or current_model in FLAGGED_MODELS
-                or row["status"] != "FINISHED"
-            )
-            if to_ignore:
-                continue
-
-            current_date = row["date"]
-            if task.benchmark == "Average":
-                current_score = np.mean(list(row["results"].values()))
-            else:
-                current_score = row["results"][task.benchmark]
-
-            if current_score > current_max:
-                if current_date == last_date and len(scores[column]) > 0:
-                    scores[column][-1] = {"model": current_model, "date": current_date, "score": current_score}
-                else:
-                    scores[column].append({"model": current_model, "date": current_date, "score": current_score})
-                current_max = current_score
-                last_date = current_date
-
-    # Step 4: Return all dictionaries as DataFrames
-    return {k: pd.DataFrame(v) for k, v in scores.items()}
-
-
-def create_plot_df(scores_df: dict[str : pd.DataFrame]) -> pd.DataFrame:
-    """
-    Transforms the scores DataFrame into a new format suitable for plotting.
-
-    :param scores_df: A DataFrame containing metric scores and dates.
-    :return: A new DataFrame reshaped for plotting purposes.
-    """
-    # Initialize the list to store DataFrames
-    dfs = []
-
-    # Iterate over the cols and create a new DataFrame for each column
-    for col in BENCHMARK_COLS + [AutoEvalColumn.average.name]:
-        d = scores_df[col].reset_index(drop=True)
-        d["task"] = col
-        dfs.append(d)
-
-    # Concatenate all the created DataFrames
-    concat_df = pd.concat(dfs, ignore_index=True)
-
-    # Sort values by 'date'
-    concat_df.sort_values(by="date", inplace=True)
-    concat_df.reset_index(drop=True, inplace=True)
-    return concat_df
-
-
-def create_metric_plot_obj(df: pd.DataFrame, metrics: list[str], title: str) -> Figure:
-    """
-    Create a Plotly figure object with lines representing different metrics
-    and horizontal dotted lines representing human baselines.
-
-    :param df: The DataFrame containing the metric values, names, and dates.
-    :param metrics: A list of strings representing the names of the metrics
-                    to be included in the plot.
-    :param title: A string representing the title of the plot.
-    :return: A Plotly figure object with lines representing metrics and
-             horizontal dotted lines representing human baselines.
-    """
-
-    # Filter the DataFrame based on the specified metrics
-    df = df[df["task"].isin(metrics)]
-
-    # Filter the human baselines based on the specified metrics
-    filtered_human_baselines = {k: v for k, v in HUMAN_BASELINE.items() if k in metrics}
-
-    # Create a line figure using plotly express with specified markers and custom data
-    fig = px.line(
-        df,
-        x="date",
-        y="score",
-        color="task",
-        markers=True,
-        custom_data=["task", "score", "model"],
-        title=title,
-    )
-
-    # Update hovertemplate for better hover interaction experience
-    fig.update_traces(
-        hovertemplate="<br>".join(
-            [
-                "Model Name: %{customdata[2]}",
-                "Metric Name: %{customdata[0]}",
-                "Date: %{x}",
-                "Metric Value: %{y}",
-            ]
-        )
-    )
-
-    # Update the range of the y-axis
-    fig.update_layout(yaxis_range=[0, 100])
-
-    # Create a dictionary to hold the color mapping for each metric
-    metric_color_mapping = {}
-
-    # Map each metric name to its color in the figure
-    for trace in fig.data:
-        metric_color_mapping[trace.name] = trace.line.color
-
-    # Iterate over filtered human baselines and add horizontal lines to the figure
-    for metric, value in filtered_human_baselines.items():
-        color = metric_color_mapping.get(metric, "blue")  # Retrieve color from mapping; default to blue if not found
-        location = "top left" if metric == "HellaSwag" else "bottom left"  # Set annotation position
-        # Add horizontal line with matched color and positioned annotation
-        fig.add_hline(
-            y=value,
-            line_dash="dot",
-            annotation_text=f"{metric} human baseline",
-            annotation_position=location,
-            annotation_font_size=10,
-            annotation_font_color=color,
-            line_color=color,
-        )
-
-    return fig
-
-
-# Example Usage:
-# human_baselines dictionary is defined.
-# chart = create_metric_plot_obj(scores_df, ["ARC", "HellaSwag", "MMLU", "TruthfulQA"], human_baselines, "Graph Title")

update_dynamic.py

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