src/backend/run_eval_suite_harness.py

import json
import logging
from datetime import datetime
from pathlib import Path
from typing import Union

from lm_eval import evaluator, utils
from lm_eval.tasks import TaskManager

from src.backend.manage_requests import EvalRequest
from src.envs import API
from src.logging import setup_logger


logging.getLogger("openai").setLevel(logging.WARNING)
logger = setup_logger(__name__)


def run_evaluation(
    eval_request: EvalRequest,
    task_names: list,
    num_fewshot: int,
    batch_size: Union[int, str],
    device: str,
    local_dir: str,
    results_repo: str,
    limit: int = None,
):
    """Runs one evaluation for the current evaluation request file, then pushes the results to the hub.

    Args:
        eval_request (EvalRequest): Input evaluation request file representation
        task_names (list): Tasks to launch
        num_fewshot (int): Number of few shots to use
        batch_size (int or str): Selected batch size or 'auto'
        device (str): "cpu" or "cuda:0", depending on what you assigned to the space
        local_dir (str): Where to save the results locally
        results_repo (str): To which repository to upload the results
        limit (int, optional): Whether to use a number of samples only for the evaluation - only for debugging

    Returns:
        _type_: _description_
    """
    if limit:
        logger.info(
            "WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT."
        )

    task_manager = TaskManager()
    all_tasks = task_manager.all_tasks
    task_names = utils.pattern_match(task_names, all_tasks)

    logger.info(f"Selected Tasks: {task_names}")

    results = evaluator.simple_evaluate(
        model="hf",
        model_args=eval_request.get_model_args(),
        tasks=task_names,
        num_fewshot=num_fewshot,
        batch_size=batch_size,
        device=device,
        limit=limit,
        write_out=True,  # Whether to write out an example document and model input, for checking task integrity
    )

    results["config"]["model_dtype"] = eval_request.precision
    results["config"]["model_name"] = eval_request.model
    results["config"]["model_sha"] = eval_request.revision

    dumped = json.dumps(results, indent=2)
    logger.info(dumped)

    results_path = Path(local_dir, eval_request.model, f"results_{datetime.now()}.json")
    results_path.parent.mkdir(exist_ok=True, parents=True)
    results_path.write_text(dumped)

    logger.info(evaluator.make_table(results))

    API.upload_file(
        path_or_fileobj=results_path,
        path_in_repo=results_path.relative_to(local_dir).as_posix(),
        repo_id=results_repo,
        repo_type="dataset",
    )

    return results