cli/halueval-cli.py

#!/usr/bin/env python

from huggingface_hub import snapshot_download

from src.backend.envs import EVAL_REQUESTS_PATH_BACKEND
from src.backend.manage_requests import get_eval_requests
from src.backend.manage_requests import EvalRequest
from src.backend.run_eval_suite import run_evaluation

from lm_eval.tasks import initialize_tasks, include_task_folder
from lm_eval import tasks, evaluator, utils

from src.backend.envs import Tasks, EVAL_REQUESTS_PATH_BACKEND, EVAL_RESULTS_PATH_BACKEND, DEVICE, LIMIT, Task
from src.envs import QUEUE_REPO


def main():
    snapshot_download(repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60)

    PENDING_STATUS = "PENDING"
    RUNNING_STATUS = "RUNNING"
    FINISHED_STATUS = "FINISHED"
    FAILED_STATUS = "FAILED"

    status = [PENDING_STATUS, RUNNING_STATUS, FINISHED_STATUS, FAILED_STATUS]

    # Get all eval request that are FINISHED, if you want to run other evals, change this parameter
    eval_requests: list[EvalRequest] = get_eval_requests(job_status=status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND)
    eval_request = [r for r in eval_requests if 'bloom-560m' in r.model][0]

    TASKS_HARNESS = [t.value for t in Tasks if 'halueval_qa' in t.value.benchmark]
    # task_names = ['triviaqa']
    # TASKS_HARNESS = [task.value for task in Tasks]

    include_task_folder("src/backend/tasks/")
    initialize_tasks('INFO')

    # breakpoint()

    print(tasks.ALL_TASKS)

    for task in TASKS_HARNESS:
        print(f"Selected Tasks: [{task}]")
        results = evaluator.simple_evaluate(model="hf", model_args=eval_request.get_model_args(), tasks=[task.benchmark], num_fewshot=0,
                                            batch_size=1, device=DEVICE, use_cache=None, limit=8, write_out=True)
        print('AAA', results["results"])

        breakpoint()


if __name__ == "__main__":
    main()