yourbench_space/app.py

import os
import sys
import time
import uuid
import asyncio
from pathlib import Path

from loguru import logger

import gradio as gr
from datasets import load_dataset
from huggingface_hub import HfApi, whoami
from yourbench_space import PATH
from yourbench_space.utils import (
    STAGES,
    SubprocessManagerGroup,
    save_files,
    update_dataset,
    map_stage_names,
    is_running_locally,
    on_generation_succsess,
)
from yourbench_space.config import generate_and_save_config
from yourbench_space.evaluation import run_evaluations, create_eval_file


project_description = """
# 🚀 YourBench
### Dynamic Benchmark Generation from Your Documents

- Create zero-shot benchmarks from your documents — no manual labeling
- Evaluate top open models and publish a leaderboard in one click
- Run locally or explore the [source on GitHub](https://github.com/huggingface/yourbench)

⚠️ **Important:** This app uses your Hugging Face token for inference and uploads — you are responsible for any usage costs

Built with 🤗 by the [Hugging Face OpenEvals team](https://huggingface.co/OpenEvals)
"""

logger.remove()
logger.add(sys.stderr, level="INFO")

# Global to store all managers per session
MANAGERS = SubprocessManagerGroup()
USER_ID_SESSION_MAP: dict[str, str] = {}


docs_path = Path(__file__).parent / "docs.md"
citation_content = (
    docs_path.read_text().split("# Citation")[-1].strip()
    if docs_path.exists()
    else "# Citation\n\nDocumentation file not found."
)


def generate_and_return(hf_org, hf_dataset_name, session_state: gr.State):
    manager = MANAGERS.get(session_state)
    if manager is None:  # should not be possible
        return (
            "❌ Config generation failed",
            gr.update(visible=False, interactive=False),
        )

    session_uid = session_state.value
    config_path = generate_and_save_config(hf_org, hf_dataset_name, session_uid, manager.config_path)
    for _ in range(5):
        time.sleep(0.5)
        if config_path.exists():
            gr.Success("✅ Config generated successfully!")
            return (
                "✅ Config saved successfully!",
                gr.update(value=str(config_path), visible=True, interactive=True),
            )

    gr.Error("Failed to generate config")
    return (
        "❌ Config generation failed",
        gr.update(visible=False, interactive=False),
    )


final_dataset = None


def update_process_status(session_state: gr.State):
    """Update process status and include exit details if process has terminated"""
    if session_state is None:
        return gr.update(value=False, label="Not running")

    manager = MANAGERS.get(session_state.value)
    if manager is None:
        return gr.update(value=False, label="Not running")

    is_running = manager.is_running()

    if not is_running:
        exit_code, exit_reason = manager.get_exit_details()
        status_text = (
            f"Process Status: Stopped - {exit_reason}, exit code - {exit_code}"
            if exit_reason
            else "Process Status: Stopped"
        )
        return gr.update(value=False, label=status_text)

    return gr.update(value=True, label="Process Status: Running")


def prepare_task(session_uid: str, oauth_token: gr.OAuthToken | None, hf_dataset_name: str, _=None):
    if oauth_token is None and not is_running_locally():
        gr.Warning("You need to log in to use this Space")
        return
    new_env = os.environ.copy()

    if oauth_token:
        new_env["HF_TOKEN"] = oauth_token.token

    new_env["DATASET_PREFIX"] = hf_dataset_name
    MANAGERS.start_process(session_uid, custom_env=new_env)


def update_hf_org_dropdown(oauth_token: gr.OAuthToken | None):
    if oauth_token is None:
        return gr.Dropdown([], label="Organization")

    try:
        user_info = whoami(oauth_token.token)
        org_names = [org["name"] for org in user_info.get("orgs", [])]
        user_name = user_info.get("name", "Unknown User")
        org_names.insert(0, user_name)
        return gr.Dropdown(org_names, value=user_name, label="Organization")

    except Exception as e:
        return gr.Dropdown([], label="Organization")


def switch_to_run_generation_tab():
    return gr.Tabs(selected=1)


def enable_button(files):
    return gr.update(interactive=bool(files))


def run_evaluation_pipeline(oauth_token: gr.OAuthToken | None, org_name, eval_name, config_name="lighteval"):
    eval_ds_name = f"{org_name}/{eval_name}"
    repo_id = f"{org_name}/leaderboard_yourbench_{eval_ds_name.replace('/', '_')}"
    folder_path = str(Path(PATH) / "yourbench_space" / "leaderboard_space")

    try:
        load_dataset(eval_ds_name, name=config_name, streaming=True, token=oauth_token.token)
    except Exception as e:
        logger.error(f"Failed to load dataset '{eval_ds_name}': {e}")
        return "❌ Failed: Dataset loading error"

    new_env = os.environ.copy()
    if oauth_token:
        new_env["HF_TOKEN"] = oauth_token.token

    try:
        create_eval_file(eval_ds_name)
        status = asyncio.run(run_evaluations(org=org_name, eval_ds_name=eval_ds_name, custom_env=new_env))
    except Exception as e:
        logger.error(f"Evaluation error: {e}")
        return f"❌ Failed: Evaluation error\n{e}"

    api = HfApi()
    space_was_regenerated = False

    try:
        api.create_repo(
            repo_id=repo_id,
            repo_type="space",
            space_sdk="gradio",
            token=oauth_token.token,
        )
    except Exception as e:
        if "409" in str(e) and "already created this space repo" in str(e):
            logger.info(f"Space '{repo_id}' already exists. Deleting and regenerating it.")
            try:
                api.delete_repo(repo_id=repo_id, repo_type="space", token=oauth_token.token)
                api.create_repo(
                    repo_id=repo_id,
                    repo_type="space",
                    space_sdk="gradio",
                    token=oauth_token.token,
                )
                space_was_regenerated = True
            except Exception as delete_err:
                logger.error(f"Failed to delete and recreate space '{repo_id}': {delete_err}")
                return f"✅ Evaluation succeeded\n❌ Failed: Could not recreate space\n{delete_err}"
        else:
            logger.error(f"Space creation error: {e}")
            return f"✅ Evaluation succeeded\n❌ Failed: Space creation error\n{e}"

    try:
        api.upload_folder(
            repo_id=repo_id,
            repo_type="space",
            folder_path=folder_path,
            token=oauth_token.token,
        )
        api.add_space_secret(
            repo_id=repo_id,
            key="HF_TOKEN",
            value=oauth_token.token,
            token=oauth_token.token,
        )
        api.add_space_variable(repo_id=repo_id, key="TASK", value=eval_ds_name, token=oauth_token.token)
        api.add_space_variable(repo_id=repo_id, key="ORG_NAME", value=org_name, token=oauth_token.token)
    except Exception as e:
        logger.error(f"Failed during space setup: {e}")
        return f"✅ Evaluation succeeded\n❌ Failed: Space setup error\n{e}"

    if space_was_regenerated:
        return f"✅ Evaluation succeeded\n🔁 Space '{repo_id}' was regenerated successfully"
    return f"✅ Evaluation and Space creation completed successfully for: {repo_id}"


def init_session(profile: gr.OAuthProfile | None):
    """Update session on load"""
    if is_running_locally():
        username = "local"
    elif profile:
        username = profile.username
    else:
        username = None

    local_uuid = USER_ID_SESSION_MAP.get(username, str(uuid.uuid4()))

    if manager := MANAGERS.get(local_uuid):
        if manager.is_running():
            logger.info(f"Found existing running session for {local_uuid}, restoring")
            return gr.State(local_uuid, delete_callback=lambda uid: MANAGERS.remove(uid))
        else:
            logger.info(f"Found existing stale session for {local_uuid}, starting new")
            MANAGERS.remove(local_uuid)
            local_uuid = str(uuid.uuid4())

    if username:
        USER_ID_SESSION_MAP[username] = local_uuid

    MANAGERS.create(local_uuid)
    logger.info(f"Started session for {local_uuid}")
    return gr.State(local_uuid, delete_callback=lambda uid: MANAGERS.remove(uid))


btn_launch_evals = gr.Button(
    "🚀 Launch Evaluation",
    visible=True,
    interactive=False,  # Start non-interactive
    variant="primary",
)

with gr.Blocks(theme=gr.themes.Default()) as app:
    session_state = gr.State()

    gr.Markdown(project_description)

    with gr.Tabs() as tabs:
        with gr.Tab("Choose Documents & Settings", id=0):
            with gr.Column():
                gr.Markdown("### 📄 Choose your documents and settings")
                gr.Markdown(
                    "Upload your source documents that will form the knowledge base for your benchmark. Set a Hugging Face organization and dataset name."
                )
                gr.Markdown(
                    "This step also generates a config file for running the benchmark pipeline. You can download it to run YourBench locally."
                )

                with gr.Row():
                    with gr.Accordion("Hugging Face Settings"):
                        login_btn = gr.LoginButton()
                        hf_org_dropdown = gr.Dropdown(choices=[], label="Organization", allow_custom_value=True)
                        app.load(update_hf_org_dropdown, inputs=None, outputs=hf_org_dropdown)

                        hf_dataset_name = gr.Textbox(
                            label="Dataset name",
                            value="yourbench",
                            info="Name of your new evaluation dataset",
                        )

                    with gr.Accordion("Upload Files"):
                        file_input = gr.File(
                            label="Upload text files",
                            file_count="multiple",
                            file_types=[".txt", ".md", ".html", ".pdf"],
                        )
                        output = gr.Textbox(label="Log")
                        file_input.upload(
                            save_files,
                            inputs=[session_state, file_input],
                            outputs=output,
                        )
                        delete_button = gr.Button("Delete Uploaded Files", visible=False)

                preview_button = gr.Button("Generate New Config", interactive=False)
                log_message = gr.Textbox(label="Log Message", visible=True)
                download_button = gr.File(label="Download Config", visible=False, interactive=False)

                file_input.change(
                    lambda files: gr.update(visible=bool(files)),
                    inputs=file_input,
                    outputs=delete_button,
                )

                file_input.change(enable_button, inputs=file_input, outputs=preview_button)

                def clean_and_confirm(uid):
                    MANAGERS.clean_workdir(uid)
                    return (
                        "🗑️ All uploaded files have been deleted!",
                        gr.update(value=None),
                        gr.update(interactive=False),
                    )

                delete_button.click(
                    clean_and_confirm,
                    inputs=session_state,
                    outputs=[output, file_input, preview_button],
                )

                preview_button.click(
                    generate_and_return,
                    inputs=[hf_org_dropdown, hf_dataset_name, session_state],
                    outputs=[log_message, download_button],
                )
                preview_button.click(
                    switch_to_run_generation_tab,
                    inputs=None,
                    outputs=tabs,
                )

        with gr.Tab("Run Benchmark Pipeline", id=1):
            with gr.Column():
                gr.Markdown("### ⚙️ Run the benchmark generation pipeline")
                gr.Markdown(
                    "Start the pipeline to process documents, generate questions, and build the private evaluation dataset. Watch logs, track progress, and preview the results."
                )

                with gr.Row():
                    start_button = gr.Button("Start Task")
                    stop_button = gr.Button("Stop Task")
                    kill_button = gr.Button("Kill Task")

                start_button.click(prepare_task, inputs=[session_state, login_btn, hf_dataset_name])
                stop_button.click(MANAGERS.stop_process, inputs=session_state)
                kill_button.click(MANAGERS.kill_process, inputs=session_state)

                process_status = gr.Checkbox(label="Process Status", interactive=False)
                status_timer = gr.Timer(2.0, active=True)
                status_timer.tick(update_process_status, inputs=session_state, outputs=process_status)

                with gr.Row():
                    with gr.Accordion("Stages", open=True):
                        stages_table = gr.CheckboxGroup(
                            choices=map_stage_names(STAGES),
                            value=[],
                            label="Pipeline Stages Completed",
                            container=False,
                            interactive=False,
                        )

                with gr.Row():
                    with gr.Column():
                        with gr.Accordion("Log Output", open=True):
                            log_output = gr.Code(language=None, lines=20, interactive=False)

                    with gr.Column():
                        with gr.Accordion("Ingestion Preview"):
                            ingestion_df = gr.DataFrame()

                        with gr.Accordion("Summarization Preview"):
                            summarization_df = gr.DataFrame()

                        with gr.Accordion("Single Shot Preview"):
                            single_shot_df = gr.DataFrame()

                        with gr.Accordion("Multi Hop Preview"):
                            multi_hop_df = gr.DataFrame()

                        with gr.Accordion("Lighteval Preview"):
                            lighteval_df = gr.DataFrame()
                    stages_table.change(
                        update_dataset,
                        inputs=[stages_table, hf_org_dropdown, hf_dataset_name],
                        outputs=[ingestion_df, summarization_df, single_shot_df, multi_hop_df, lighteval_df],
                    )

                    stages_table.change(
                        on_generation_succsess,
                        inputs=stages_table,
                        outputs=[tabs, btn_launch_evals],
                    )

                    # TODO: this timer should only be active when the second tab is passed to active for the first time
                    log_timer = gr.Timer(1.0, active=True)
                    log_timer.tick(
                        MANAGERS.read_and_get_output,
                        inputs=session_state,
                        outputs=[log_output, stages_table],
                    )

        with gr.Tab("Evaluate Models on Benchmark", id=2):
            with gr.Column():
                gr.Markdown("### 🧪 Evaluate models on your benchmark")
                gr.Markdown(
                    "Runs the evaluation with [Lighteval](https://github.com/huggingface/lighteval) on the resulted dataset using 5+ open models, then deploys a leaderboard as a Hugging Face Space under your org."
                )

                with gr.Row():
                    with gr.Column():
                        btn_launch_evals.render()
                    with gr.Column():
                        clear_status_btn = gr.Button("Clear", variant="secondary")

                with gr.Accordion("Evaluation Log", open=True):
                    eval_status = gr.Textbox(label="", lines=6, interactive=False, show_label=False)

                btn_launch_evals.click(
                    run_evaluation_pipeline,
                    [hf_org_dropdown, hf_dataset_name, gr.State("lighteval")],
                    eval_status,
                )
                clear_status_btn.click(lambda: "", outputs=eval_status)

    app.load(init_session, outputs=session_state)

app.launch(allowed_paths=[PATH])