src/backend/run_eval_suite.py

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

from src.backend.manage_requests import EvalRequest

from src.backend.tasks.xsum.task import XSum
from src.backend.tasks.cnndm.task import CNNDM
from src.backend.tasks.selfcheckgpt.task import SelfCheckGpt


def run_evaluation(eval_request: EvalRequest, task_names, num_fewshot, batch_size, device, use_cache=None, limit=None, max_nb_samples=100) -> dict:
    if limit:
        print("WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT.")

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

    print(f"Considered Tasks: {task_names}")
    print(f"Allowed Tasks: {tasks.ALL_TASKS}")

    task_names = utils.pattern_match(task_names, tasks.ALL_TASKS)

    print(f"Selected Tasks: {task_names}")
    print(f"Eval Request: {eval_request.get_model_args()}")

    results = evaluator.simple_evaluate(model="hf-auto",  # "hf-causal-experimental",  # "hf-causal"
                                        model_args=eval_request.get_model_args(),
                                        tasks=task_names, num_fewshot=num_fewshot,
                                        batch_size=batch_size, device=device, use_cache=use_cache,
                                        limit=limit, write_out=True)

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

    if max_nb_samples is not None:
        if 'samples' in results:
            samples = results['samples']
            for task_name in samples.keys():
                if len(samples[task_name]) > max_nb_samples:
                    results['samples'][task_name] = results['samples'][task_name][:max_nb_samples]

    # print(evaluator.make_table(results))

    return results