app.py

import os
import shutil
import numpy as np
import gradio as gr
from huggingface_hub import Repository, HfApi
from transformers import AutoConfig
import json
from apscheduler.schedulers.background import BackgroundScheduler
import pandas as pd
import datetime
from utils import get_eval_results_dicts, make_clickable_model

# clone / pull the lmeh eval data
H4_TOKEN = os.environ.get("H4_TOKEN", None)
LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", None))


def get_all_requested_models(requested_models_dir):
    depth = 1
    file_names = []

    for root, dirs, files in os.walk(requested_models_dir):
        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
        if current_depth == depth:
            file_names.extend([os.path.join(root, file) for file in files])

    return set([file_name.lower().split("./evals/")[1] for file_name in file_names])


repo = None
requested_models = None
if H4_TOKEN:
    print("pulling repo")
    # try:
    #     shutil.rmtree("./evals/")
    # except:
    #     pass

    repo = Repository(
        local_dir="./evals/",
        clone_from=LMEH_REPO,
        use_auth_token=H4_TOKEN,
        repo_type="dataset",
    )
    repo.git_pull()

    requested_models_dir = "./evals/eval_requests"
    requested_models = get_all_requested_models(requested_models_dir)


# parse the results
BENCHMARKS = ["arc_challenge", "hellaswag", "hendrycks", "truthfulqa_mc"]

METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]


def load_results(model, benchmark, metric):
    file_path = os.path.join("evals", model, f"{model}-eval_{benchmark}.json")
    if not os.path.exists(file_path):
        return 0.0, None

    with open(file_path) as fp:
        data = json.load(fp)
    accs = np.array([v[metric] for k, v in data["results"].items()])
    mean_acc = np.mean(accs)
    return mean_acc, data["config"]["model_args"]


COLS = [
    "Model",
    "Revision",
    "Average ⬆️",
    "ARC (25-shot) ⬆️",
    "HellaSwag (10-shot) ⬆️",
    "MMLU (5-shot) ⬆️",
    "TruthfulQA (0-shot) ⬆️",
]
TYPES = [
    "markdown",
    "str",
    "number",
    "number",
    "number",
    "number",
    "number",
]

if not IS_PUBLIC:
    COLS.insert(2, "8bit")
    TYPES.insert(2, "bool")

EVAL_COLS = ["model", "revision", "private", "8bit_eval", "is_delta_weight", "status"]
EVAL_TYPES = ["markdown", "str", "bool", "bool", "bool", "str"]

BENCHMARK_COLS = [
    "ARC (25-shot) ⬆️",
    "HellaSwag (10-shot) ⬆️",
    "MMLU (5-shot) ⬆️",
    "TruthfulQA (0-shot) ⬆️",
]


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():
    if repo:
        print("pulling changes")
        repo.git_pull()

    all_data = get_eval_results_dicts(IS_PUBLIC)

    if not IS_PUBLIC:
        gpt4_values = {
            "Model": f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt4</a>',
            "Revision": "tech report",
            "8bit": None,
            "Average ⬆️": 84.3,
            "ARC (25-shot) ⬆️": 96.3,
            "HellaSwag (10-shot) ⬆️": 95.3,
            "MMLU (5-shot) ⬆️": 86.4,
            "TruthfulQA (0-shot) ⬆️": 59.0,
        }
        all_data.append(gpt4_values)
        gpt35_values = {
            "Model": f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt3.5</a>',
            "Revision": "tech report",
            "8bit": None,
            "Average ⬆️": 71.9,
            "ARC (25-shot) ⬆️": 85.2,
            "HellaSwag (10-shot) ⬆️": 85.5,
            "MMLU (5-shot) ⬆️": 70.0,
            "TruthfulQA (0-shot) ⬆️": 47.0,
        }
        all_data.append(gpt35_values)

    df = pd.DataFrame.from_records(all_data)
    df = df.sort_values(by=["Average ⬆️"], ascending=False)
    df = df[COLS]

    # get incomplete models
    incomplete_models = df[has_nan_values(df, BENCHMARK_COLS)]["Model"].tolist()
    print(
        [
            model.split(" style")[0].split("https://huggingface.co/")[1]
            for model in incomplete_models
        ]
    )

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


def get_eval_table():
    if repo:
        print("pulling changes for eval")
        repo.git_pull()
    entries = [
        entry
        for entry in os.listdir("evals/eval_requests")
        if not entry.startswith(".")
    ]
    all_evals = []

    for entry in entries:
        # print(entry)
        if ".json" in entry:
            file_path = os.path.join("evals/eval_requests", 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)
        else:
            # this is a folder
            sub_entries = [
                e
                for e in os.listdir(f"evals/eval_requests/{entry}")
                if not e.startswith(".")
            ]
            for sub_entry in sub_entries:
                file_path = os.path.join("evals/eval_requests", 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"] == "PENDING"]
    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
    finished_list = [e for e in all_evals if e["status"] == "FINISHED"]
    df_pending = pd.DataFrame.from_records(pending_list)
    df_running = pd.DataFrame.from_records(running_list)
    df_finished = pd.DataFrame.from_records(finished_list)
    return df_finished[EVAL_COLS], df_running[EVAL_COLS], df_pending[EVAL_COLS]


leaderboard = get_leaderboard()
finished_eval_queue, running_eval_queue, pending_eval_queue = get_eval_table()


def is_model_on_hub(model_name, revision) -> bool:
    try:
        config = AutoConfig.from_pretrained(model_name, revision=revision)
        return True

    except Exception as e:
        print("Could not get the model config from the hub")
        print(e)
        return False


def add_new_eval(
    model: str,
    base_model: str,
    revision: str,
    is_8_bit_eval: bool,
    private: bool,
    is_delta_weight: bool,
):
    # check the model actually exists before adding the eval
    if revision == "":
        revision = "main"
    if is_delta_weight and not is_model_on_hub(base_model, revision):
        error_message = f'Base model "{base_model}" was not found on hub!'
        print(error_message)
        return f"<p style='color: red; font-size: 20px; text-align: center;'>{error_message}</p>"

    if not is_model_on_hub(model, revision):
        error_message = f'Model "{model}"was not found on hub!'
        print(error_message)
        return f"<p style='color: red; font-size: 20px; text-align: center;'>{error_message}</p>"

    print("adding new eval")

    eval_entry = {
        "model": model,
        "base_model": base_model,
        "revision": revision,
        "private": private,
        "8bit_eval": is_8_bit_eval,
        "is_delta_weight": is_delta_weight,
        "status": "PENDING",
    }

    user_name = ""
    model_path = model
    if "/" in model:
        user_name = model.split("/")[0]
        model_path = model.split("/")[1]

    OUT_DIR = f"eval_requests/{user_name}"
    os.makedirs(OUT_DIR, exist_ok=True)
    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{is_8_bit_eval}_{is_delta_weight}.json"

    # Check for duplicate submission
    if out_path.lower() in requested_models:
        duplicate_request_message = "This model has been already submitted."
        return f"<p style='color: orange; font-size: 20px; text-align: center;'>{duplicate_request_message}</p>"

    with open(out_path, "w") as f:
        f.write(json.dumps(eval_entry))
    LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"

    api = HfApi()
    api.upload_file(
        path_or_fileobj=out_path,
        path_in_repo=out_path,
        repo_id=LMEH_REPO,
        token=H4_TOKEN,
        repo_type="dataset",
    )

    success_message = "Your request has been submitted to the evaluation queue!"
    return f"<p style='color: green; font-size: 20px; text-align: center;'>{success_message}</p>"


def refresh():
    leaderboard = get_leaderboard()
    finished_eval_queue, running_eval_queue, pending_eval_queue = get_eval_table()
    get_leaderboard(), get_eval_table()
    return leaderboard, finished_eval_queue, running_eval_queue, pending_eval_queue


block = gr.Blocks()
with block:
    with gr.Row():
        gr.Markdown(
            f"""
# 🤗 Open LLM Leaderboard
<font size="4">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. The 🤗 Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. We evaluate models on 4 key benchmarks from 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 key advantage of this leaderboard is that anyone from the community can submit a model for automated evaluation on the 🤗 GPU cluster, as long as it is a 🤗 Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as LLaMa.

Evaluation is performed against 4 popular benchmarks:
- <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 benchmark to measure whether a language model is truthful in generating answers to questions.

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. </font>
        """
        )

    with gr.Row():
        leaderboard_table = gr.components.Dataframe(
            value=leaderboard, headers=COLS, datatype=TYPES, max_rows=5
        )

    with gr.Row():
        gr.Markdown(
            f"""
    # Evaluation Queue for the 🤗 Open LLM Leaderboard, these models will be automatically evaluated on the 🤗 cluster

    """
        )
    with gr.Accordion("Finished Evaluations", open=False):
        with gr.Row():
            finished_eval_table = gr.components.Dataframe(
                value=finished_eval_queue,
                headers=EVAL_COLS,
                datatype=EVAL_TYPES,
                max_rows=5,
            )
    with gr.Accordion("Running Evaluation Queue", open=False):
        with gr.Row():
            running_eval_table = gr.components.Dataframe(
                value=running_eval_queue,
                headers=EVAL_COLS,
                datatype=EVAL_TYPES,
                max_rows=5,
            )

    with gr.Accordion("Running & Pending Evaluation Queue", open=False):
        with gr.Row():
            pending_eval_table = gr.components.Dataframe(
                value=pending_eval_queue,
                headers=EVAL_COLS,
                datatype=EVAL_TYPES,
                max_rows=5,
            )

    with gr.Row():
        refresh_button = gr.Button("Refresh")
        refresh_button.click(
            refresh,
            inputs=[],
            outputs=[
                leaderboard_table,
                finished_eval_table,
                running_eval_table,
                pending_eval_table,
            ],
        )

    with gr.Accordion("Submit a new model for evaluation"):
        with gr.Row():
            with gr.Column():
                model_name_textbox = gr.Textbox(label="Model name")
                revision_name_textbox = gr.Textbox(label="revision", placeholder="main")

            with gr.Column():
                is_8bit_toggle = gr.Checkbox(
                    False, label="8 bit eval", visible=not IS_PUBLIC
                )
                private = gr.Checkbox(False, label="Private", visible=not IS_PUBLIC)
                is_delta_weight = gr.Checkbox(False, label="Delta weights")
                base_model_name_textbox = gr.Textbox(label="base model (for delta)")

        with gr.Row():
            submit_button = gr.Button("Submit Eval")
        with gr.Row():
            submission_result = gr.Markdown()
            submit_button.click(
                add_new_eval,
                [
                    model_name_textbox,
                    base_model_name_textbox,
                    revision_name_textbox,
                    is_8bit_toggle,
                    private,
                    is_delta_weight,
                ],
                submission_result,
            )

    block.load(
        refresh,
        inputs=[],
        outputs=[
            leaderboard_table,
            finished_eval_table,
            running_eval_table,
            pending_eval_table,
        ],
    )
block.launch()