rodrigomasini/data_only_hallucination_leaderboard

1#!/usr/bin/env python2 3import json4import os5import time6 7from datetime import datetime, timezone8 9from src.envs import API, EVAL_REQUESTS_PATH, H4_TOKEN, QUEUE_REPO10from src.submission.check_validity import already_submitted_models, get_model_size, is_model_on_hub11 12from huggingface_hub import snapshot_download13from src.backend.envs import EVAL_REQUESTS_PATH_BACKEND14from src.backend.manage_requests import get_eval_requests15from src.backend.manage_requests import EvalRequest16 17 18def add_new_eval(model: str, base_model: str, revision: str, precision: str, private: bool, weight_type: str, model_type: str):19    REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)20 21    user_name = ""22    model_path = model23    if "/" in model:24        tokens = model.split("/")25        user_name = tokens[0]26        model_path = tokens[1]27 28    precision = precision.split(" ")[0]29    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")30 31    if model_type is None or model_type == "":32        return print("Please select a model type.")33 34    # Does the model actually exist?35    if revision == "":36        revision = "main"37 38    # Is the model on the hub?39    if weight_type in ["Delta", "Adapter"]:40        base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=H4_TOKEN, test_tokenizer=True)41        if not base_model_on_hub:42            print(f'Base model "{base_model}" {error}')43            return44 45    if not weight_type == "Adapter":46        model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, test_tokenizer=True)47        if not model_on_hub:48            print(f'Model "{model}" {error}')49            return50 51    # Is the model info correctly filled?52    try:53        model_info = API.model_info(repo_id=model, revision=revision)54    except Exception:55        print("Could not get your model information. Please fill it up properly.")56        return57 58    model_size = get_model_size(model_info=model_info, precision=precision)59 60    license = 'none'61    try:62        license = model_info.cardData["license"]63    except Exception:64        print("Please select a license for your model")65        # return66 67    # modelcard_OK, error_msg = check_model_card(model)68    # if not modelcard_OK:69    #     print(error_msg)70    #     return71 72    # Seems good, creating the eval73    print("Adding new eval")74 75    eval_entry = {76        "model": model,77        "base_model": base_model,78        "revision": revision,79        "private": private,80        "precision": precision,81        "weight_type": weight_type,82        "status": "PENDING",83        "submitted_time": current_time,84        "model_type": model_type,85        "likes": model_info.likes,86        "params": model_size,87        "license": license,88    }89 90    # Check for duplicate submission91    if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:92        print("This model has been already submitted.")93        return94 95    print("Creating eval file")96    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"97    os.makedirs(OUT_DIR, exist_ok=True)98    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json"99 100    with open(out_path, "w") as f:101        f.write(json.dumps(eval_entry))102 103    print("Uploading eval file")104    API.upload_file(path_or_fileobj=out_path, path_in_repo=out_path.split("eval-queue/")[1],105                    repo_id=QUEUE_REPO, repo_type="dataset", commit_message=f"Add {model} to eval queue")106 107    # Remove the local file108    os.remove(out_path)109 110    print("Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list.")111    return112 113 114def main():115    from huggingface_hub import HfApi116 117    api = HfApi()118    model_lst = api.list_models()119 120    model_lst = [m for m in model_lst]121 122    def custom_filter(m) -> bool:123        # res = m.pipeline_tag in {'text-generation'} and 'en' in m.tags and m.private is False124        # res = m.pipeline_tag in {'text-generation'} and 'en' in m.tags and m.private is False and 'mistralai/' in m.id125        res = 'mistralai/' in m.id126        return res127 128    filtered_model_lst = sorted([m for m in model_lst if custom_filter(m)], key=lambda m: m.downloads, reverse=True)129 130    snapshot_download(repo_id=QUEUE_REPO, revision="main", local_dir=EVAL_REQUESTS_PATH_BACKEND, repo_type="dataset", max_workers=60)131 132    PENDING_STATUS = "PENDING"133    RUNNING_STATUS = "RUNNING"134    FINISHED_STATUS = "FINISHED"135    FAILED_STATUS = "FAILED"136 137    status = [PENDING_STATUS, RUNNING_STATUS, FINISHED_STATUS, FAILED_STATUS]138 139    # Get all eval requests140    eval_requests: list[EvalRequest] = get_eval_requests(job_status=status, hf_repo=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH_BACKEND)141 142    requested_model_names = {e.model for e in eval_requests}143 144    # breakpoint()145 146    for i in range(min(200, len(filtered_model_lst))):147        model = filtered_model_lst[i]148 149        print(f'Considering {model.id} ..')150 151        is_finetuned = any(tag.startswith('base_model:') for tag in model.tags)152 153        model_type = 'pretrained'154        if is_finetuned:155            model_type = "fine-tuned"156 157        is_instruction_tuned = 'nstruct' in model.id158        if is_instruction_tuned:159            model_type = "instruction-tuned"160 161        if model.id not in requested_model_names:162 163            if 'mage' not in model.id:164                add_new_eval(model=model.id, base_model='', revision='main', precision='float32', private=False, weight_type='Original', model_type=model_type)165                time.sleep(10)166        else:167            print(f'Model {model.id} already added, not adding it to the queue again.')168 169 170if __name__ == "__main__":171    main()172