app.py

import json
import os
from datetime import datetime, timezone


import gradio as gr
import numpy as np
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import HfApi
from transformers import AutoConfig

from src.auto_leaderboard.get_model_metadata import apply_metadata
from src.assets.text_content import *
from src.auto_leaderboard.load_results import get_eval_results_dicts, make_clickable_model
from src.assets.hardcoded_evals import gpt4_values, gpt35_values, baseline
from src.assets.css_html_js import custom_css, get_window_url_params
from src.utils_display import AutoEvalColumn, EvalQueueColumn, fields, styled_error, styled_warning, styled_message
from src.init import get_all_requested_models, load_all_info_from_hub

# 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"

IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))

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()

def restart_space():
    api.restart_space(
        repo_id="HuggingFaceH4/open_llm_leaderboard", token=H4_TOKEN
    )

eval_queue, requested_models, eval_results = 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

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]

if not IS_PUBLIC:
    COLS.insert(2, AutoEvalColumn.precision.name)
    TYPES.insert(2, AutoEvalColumn.precision.type)

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]]


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)


def get_leaderboard_df_1():
    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(IS_PUBLIC)

    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]

    # filter out if any of the benchmarks have not been produced
    df = df[has_no_nan_values(df, BENCHMARK_COLS)]
    print(df)
    print(type(df))
    return df

def get_leaderboard_df():
    
    data = {
        'Datasets': ['SOTA(FT)', 'SOTA(ZS)', 'FLAN-T5', 'GPT-3', 'GPT-3.5v2', 'GPT-3.5v3', 'ChatGPT', 'GPT-4'],
        'KQApro': [93.85, 94.20, 37.27, 38.28, 38.01, 40.35, 47.93, 57.20],
        'LC-quad2': [33.10, '-', 30.14, 33.04, 33.77, 39.04, 42.76, 54.95],
        'WQSP': [73.10, 62.98, 59.87, 67.68, 72.34, 79.60, 83.70, 90.45],
        'CWQ': [72.20, '-', 46.69, 51.77, 53.96, 57.54, 64.02, 71.00],
        'GrailQA': [76.31, '-', 29.02, 27.58, 30.50, 35.43, 46.77, 51.40],
        'GraphQ': [41.30, '-', 32.27, 38.32, 40.85, 47.95, 53.10, 63.20],
        'QALD-9': [67.82, '-', 30.17, 38.54, 44.96, 46.19, 45.71, 57.20],
        'MKQA': [46.00, '-', 20.17, 26.97, 30.14, 39.05, 44.30, 59.20]
    }
    
    df = pd.DataFrame(data)
    return df

def get_evaluation_queue_df():
    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(EVAL_REQUESTS_PATH)
        if not entry.startswith(".")
    ]
    all_evals = []

    for entry in entries:
        if ".json" in entry:
            file_path = os.path.join(EVAL_REQUESTS_PATH, entry)
            with open(file_path) as fp:
                data = json.load(fp)

            data["# params"] = "unknown"
            data["model"] = make_clickable_model(data["model"])
            data["revision"] = 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"{EVAL_REQUESTS_PATH}/{entry}")
                if not e.startswith(".")
            ]
            for sub_entry in sub_entries:
                file_path = os.path.join(EVAL_REQUESTS_PATH, entry, sub_entry)
                with open(file_path) as fp:
                    data = json.load(fp)

                # data["# params"] = get_n_params(data["model"])
                data["model"] = make_clickable_model(data["model"])
                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")]
    df_pending = pd.DataFrame.from_records(pending_list, columns=EVAL_COLS)
    df_running = pd.DataFrame.from_records(running_list, columns=EVAL_COLS)
    df_finished = pd.DataFrame.from_records(finished_list, columns=EVAL_COLS)
    return df_finished[EVAL_COLS], df_running[EVAL_COLS], df_pending[EVAL_COLS]



original_df = get_leaderboard_df()
leaderboard_df = original_df.copy()
(
    finished_eval_queue_df,
    running_eval_queue_df,
    pending_eval_queue_df,
) = get_evaluation_queue_df()

def is_model_on_hub(model_name, revision) -> bool:
    try:
        AutoConfig.from_pretrained(model_name, revision=revision)
        return True, None
    
    except ValueError as e:
        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!"


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")

    # check the model actually exists before adding the eval
    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}')
    
    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]

    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 out_path.split("eval-queue/")[1].lower() 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))

    api.upload_file(
        path_or_fileobj=out_path,
        path_in_repo=out_path.split("eval-queue/")[1],
        repo_id=QUEUE_REPO,
        token=H4_TOKEN,
        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.")


def refresh():
    leaderboard_df = get_leaderboard_df()
    (
        finished_eval_queue_df,
        running_eval_queue_df,
        pending_eval_queue_df,
    ) = get_evaluation_queue_df()
    return (
        leaderboard_df,
        finished_eval_queue_df,
        running_eval_queue_df,
        pending_eval_queue_df,
    )


def search_table(df, query):
    if AutoEvalColumn.model_type.name in df.columns:
        filtered_df = df[
            (df[AutoEvalColumn.dummy.name].str.contains(query, case=False))
            | (df[AutoEvalColumn.model_type.name].str.contains(query, case=False))
            ]
    else:
        filtered_df = df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))]
    return filtered_df


def change_tab(query_param):
    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)


demo = gr.Blocks(css=custom_css)
with demo:
    gr.HTML(TITLE)
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
    with gr.Row():
        with gr.Box(elem_id="search-bar-table-box"):
            search_bar = gr.Textbox(
                placeholder="🔍 Search your model and press ENTER...",
                show_label=False,
                elem_id="search-bar",
            )

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        
        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=1):
            leaderboard_table = gr.components.Dataframe(
                value=leaderboard_df,
                headers=COLS,
                datatype=TYPES,
                max_rows=None,
                elem_id="leaderboard-table",
            )

            # Dummy leaderboard for handling the case when the user uses backspace key
            hidden_leaderboard_table_for_search = gr.components.Dataframe(
                value=original_df,
                headers=COLS,
                datatype=TYPES,
                max_rows=None,
                visible=False,
            )
            search_bar.submit(
                search_table,
                [hidden_leaderboard_table_for_search, search_bar],
                leaderboard_table,
            )
        with gr.TabItem("About", elem_id="llm-benchmark-tab-table", id=2):
            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")

        

    with gr.Row():
        with gr.Accordion("📙 Citation", open=False):
            citation_button = gr.Textbox(
                value=CITATION_BUTTON_TEXT,
                label=CITATION_BUTTON_LABEL,
                elem_id="citation-button",
            ).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", seconds=3600)
scheduler.start()
demo.queue(concurrency_count=40).launch()