import json import os from datetime import datetime, timezone from src.display.formatting import styled_error, styled_message, styled_warning from src.envs import API, EVAL_REQUESTS_PATH, H4_TOKEN, QUEUE_REPO, RATE_LIMIT_PERIOD, RATE_LIMIT_QUOTA from src.leaderboard.filter_models import DO_NOT_SUBMIT_MODELS from src.submission.check_validity import ( already_submitted_models, check_model_card, get_model_size, is_model_on_hub, user_submission_permission, ) ## it just uploads request file. where does the evaluation actually happen? REQUESTED_MODELS = None USERS_TO_SUBMISSION_DATES = None def add_new_eval( model: str, requested_tasks: list, # write better type hints. this is list of class Task. base_model: str, revision: str, precision: str, private: bool, weight_type: str, model_type: str, ): global REQUESTED_MODELS global USERS_TO_SUBMISSION_DATES if not REQUESTED_MODELS: REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH) # REQUESTED_MODELS is set(file_names), where file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}") user_name = "" model_path = model if "/" in model: user_name = model.split("/")[0] model_path = model.split("/")[1] precision = precision.split(" ")[0] current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ") if model_type is None or model_type == "": return styled_error("Please select a model type.") # Is the user rate limited? if user_name != "": user_can_submit, error_msg = user_submission_permission( user_name, USERS_TO_SUBMISSION_DATES, RATE_LIMIT_PERIOD, RATE_LIMIT_QUOTA ) if not user_can_submit: return styled_error(error_msg) # Did the model authors forbid its submission to the leaderboard? if model in DO_NOT_SUBMIT_MODELS or base_model in DO_NOT_SUBMIT_MODELS: return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.") # Does the model actually exist? if revision == "": revision = "main" # Is the model on the hub? if weight_type in ["Delta", "Adapter"]: base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=H4_TOKEN, test_tokenizer=True) if not base_model_on_hub: return styled_error(f'Base model "{base_model}" {error}') if not weight_type == "Adapter": model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, test_tokenizer=True) if not model_on_hub: return styled_error(f'Model "{model}" {error}') # Is the model info correctly filled? try: model_info = API.model_info(repo_id=model, revision=revision) except Exception: return styled_error("Could not get your model information. Please fill it up properly.") model_size = get_model_size(model_info=model_info, precision=precision) # Were the model card and license filled? try: license = model_info.cardData["license"] except Exception: return styled_error("Please select a license for your model") modelcard_OK, error_msg = check_model_card(model) if not modelcard_OK: return styled_error(error_msg) # Seems good, creating the eval print("Adding new eval") print() print(f"requested_tasks: {requested_tasks}") print(f"type(requested_tasks): {type(requested_tasks)}") print() # requested_tasks: [{'benchmark': 'hellaswag', 'metric': 'acc_norm', 'col_name': 'HellaSwag'}, {'benchmark': 'pubmedqa', 'metric': 'acc', 'col_name': 'PubMedQA'}] # type(requested_tasks): requested_task_names = [task_dic['benchmark'] for task_dic in requested_tasks] print() print(f"requested_task_names: {requested_task_names}") print(f"type(requested_task_names): {type(requested_task_names)}") print() already_submitted_tasks = [] for requested_task_name in requested_task_names: if f"{model}_{requested_task_name}_{revision}_{precision}" in REQUESTED_MODELS: # return styled_warning("This model has been already submitted.") already_submitted_tasks.append(requested_task_name) task_names_for_eval = set(requested_task_names) - set(already_submitted_tasks) task_names_for_eval = list(task_names_for_eval) return_msg = "Your request has been submitted to the evaluation queue! Please wait for up to an hour for the model to show in the PENDING list." if len(already_submitted_tasks) > 0: return_msg = f"This model has been already submitted for task(s) {already_submitted_tasks}. Evaluation will proceed for tasks {task_names_for_eval}. Please wait for up to an hour for the model to show in the PENDING list." if len(task_names_for_eval)==0: return styled_warning(f"This model has been already submitted for task(s) {already_submitted_tasks}.") tasks_for_eval = [dct for dct in requested_tasks if dct['benchmark'] in task_names_for_eval] print() print(f"tasks_for_eval: {tasks_for_eval}") # print(f"type(requested_task_names): {type(requested_task_names)}") print() eval_entry = { "model": model, "requested_tasks": tasks_for_eval, # this is a list of tasks. would eval file be written correctly for each tasks? YES. run_evaluation() takes list of tasks. might have to specify "base_model": base_model, "revision": revision, "private": private, "precision": precision, "weight_type": weight_type, "status": "PENDING", "submitted_time": current_time, "model_type": model_type, "likes": model_info.likes, "params": model_size, "license": license, } ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- ####---- print("Creating eval file") OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}" # local path os.makedirs(OUT_DIR, exist_ok=True) out_path = f"{OUT_DIR}/{model_path}_{'_'.join([f'{task}' for task in task_names_for_eval])}_eval_request_{private}_{precision}_{weight_type}.json" print(f"out_path = {out_path}") with open(out_path, "w") as f: f.write(json.dumps(eval_entry)) # local path used! for saving request file. print("Uploading eval file (QUEUE_REPO)") print() print(f"path_or_fileobj={out_path}, path_in_repo={out_path.split('eval-queue/')[1]}, repo_id={QUEUE_REPO}, repo_type=dataset,") API.upload_file( path_or_fileobj=out_path, path_in_repo=out_path.split("eval-queue/")[1], repo_id=QUEUE_REPO, repo_type="dataset", commit_message=f"Add {model} to eval queue", ) print(f"is os.remove(out_path) the problem?") # Remove the local file os.remove(out_path) return styled_message( return_msg )