import json from pathlib import Path from json import JSONDecodeError import logging import math from dataclasses import dataclass, field from typing import Optional, Dict, List from tqdm import tqdm from tqdm.contrib.logging import logging_redirect_tqdm import numpy as np from src.display.formatting import make_clickable_model from src.display.utils import AutoEvalColumn, ModelType, Precision, Tasks, WeightType, parse_datetime # Configure logging logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') @dataclass class EvalResult: # Also see src.display.utils.AutoEvalColumn for what will be displayed. eval_name: str # org_model_precision (uid) full_model: str # org/model (path on hub) org: Optional[str] model: str revision: str # commit hash, "" if main results: Dict[str, float] precision: Precision = Precision.Unknown model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ... weight_type: WeightType = WeightType.Original architecture: str = "Unknown" # From config file license: str = "?" likes: int = 0 num_params: int = 0 date: str = "" # submission date of request file still_on_hub: bool = True is_merge: bool = False not_flagged: bool = False status: str = "FINISHED" # List of tags, initialized to a new empty list for each instance to avoid the pitfalls of mutable default arguments. tags: List[str] = field(default_factory=list) @classmethod def init_from_json_file(cls, json_filepath: str) -> 'EvalResult': with open(json_filepath, 'r') as fp: data = json.load(fp) config = data.get("config_general", {}) precision = Precision.from_str(config.get("model_dtype", "unknown")) org_and_model = config.get("model_name", "").split("/", 1) org = org_and_model[0] if len(org_and_model) > 1 else None model = org_and_model[-1] if len(org_and_model) == 1: org = None model = org_and_model[0] result_key = f"{model}_{precision.value.name}" else: org = org_and_model[0] model = org_and_model[1] result_key = f"{org}_{model}_{precision.value.name}" full_model = "/".join(org_and_model) results = cls.extract_results(data) # Properly call the method to extract results return cls( eval_name=result_key, full_model=full_model, org=org, model=model, results=results, precision=precision, revision=config.get("model_sha", "") ) @staticmethod def extract_results(data: Dict) -> Dict[str, float]: """ Extract and process benchmark results from a given dict. Parameters: - data (Dict): A dictionary containing benchmark data. This dictionary must include 'versions' and 'results' keys with respective sub-data. Returns: - Dict[str, float]: A dictionary where keys are benchmark names and values are the processed average scores as percentages. Notes: - The method specifically checks for certain benchmark names to skip outdated entries. - Handles NaN values by setting the corresponding benchmark result to 0.0. - Averages scores across metrics for benchmarks found in the data, in a percentage format. """ results = {} for task in Tasks: task = task.value # We skip old mmlu entries if task.benchmark == "hendrycksTest": for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]: if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0: continue # Some benchamrk values are NaNs, mostly truthfulQA # Would be more optimal (without the whole dict itertion) if benchmark name was same as key in results # e.g. not harness|truthfulqa:mc|0 but truthfulqa:mc for k, v in data["results"].items(): if task.benchmark in k: if math.isnan(float(v[task.metric])): results[task.benchmark] = 0.0 continue # We average all scores of a given metric (mostly for mmlu) accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k]) if accs.size == 0 or any([acc is None for acc in accs]): continue mean_acc = np.mean(accs) * 100.0 results[task.benchmark] = mean_acc return results def update_with_request_file(self, requests_path): """Finds the relevant request file for the current model and updates info with it.""" try: request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name) if request_file is None: logging.warning(f"No request file for {self.org}/{self.model}") self.status = "FAILED" return with open(request_file, "r") as f: request = json.load(f) self.model_type = ModelType.from_str(request.get("model_type", "Unknown")) self.weight_type = WeightType[request.get("weight_type", "Original")] self.num_params = int(request.get("params", 0)) # Ensuring type safety self.date = request.get("submitted_time", "") self.architecture = request.get("architectures", "Unknown") self.status = request.get("status", "FAILED") except FileNotFoundError: self.status = "FAILED" logging.error(f"Request file: {request_file} not found for {self.org}/{self.model}") except JSONDecodeError: self.status = "FAILED" logging.error(f"Error decoding JSON from the request file for {self.org}/{self.model}") except KeyError as e: self.status = "FAILED" logging.error(f"Key error {e} in processing request file for {self.org}/{self.model}") except Exception as e: # Catch-all for any other unexpected exceptions self.status = "FAILED" logging.error(f"Unexpected error {e} for {self.org}/{self.model}") def update_with_dynamic_file_dict(self, file_dict): """Update object attributes based on the provided dictionary, with error handling for missing keys and type validation.""" # Default values set for optional or potentially missing keys. self.license = file_dict.get("license", "?") self.likes = int(file_dict.get("likes", 0)) # Ensure likes is treated as an integer self.still_on_hub = file_dict.get("still_on_hub", False) # Default to False if key is missing self.tags = file_dict.get("tags", []) # Calculate `flagged` only if 'tags' is not empty and avoid calculating each time self.not_flagged = not (any("flagged" in tag for tag in self.tags)) def to_dict(self): """Converts the Eval Result to a dict compatible with our dataframe display""" average = sum([v for v in self.results.values() if v is not None]) / len(Tasks) data_dict = { "eval_name": self.eval_name, # not a column, just a save name, AutoEvalColumn.precision.name: self.precision.value.name, AutoEvalColumn.model_type.name: self.model_type.value.name, AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol, AutoEvalColumn.weight_type.name: self.weight_type.value.name, AutoEvalColumn.architecture.name: self.architecture, AutoEvalColumn.model.name: make_clickable_model(self.full_model), AutoEvalColumn.fullname.name: self.full_model, AutoEvalColumn.revision.name: self.revision, AutoEvalColumn.average.name: average, AutoEvalColumn.license.name: self.license, AutoEvalColumn.likes.name: self.likes, AutoEvalColumn.params.name: self.num_params, AutoEvalColumn.still_on_hub.name: self.still_on_hub, AutoEvalColumn.merged.name: not( "merge" in self.tags if self.tags else False), AutoEvalColumn.moe.name: not ( ("moe" in self.tags if self.tags else False) or "moe" in self.full_model.lower()) , AutoEvalColumn.not_flagged.name: self.not_flagged, } for task in Tasks: data_dict[task.value.col_name] = self.results[task.value.benchmark] return data_dict def get_request_file_for_model(requests_path, model_name, precision): """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED""" requests_path = Path(requests_path) pattern = f"{model_name}_eval_request_*.json" # Using pathlib to find files matching the pattern request_files = list(requests_path.glob(pattern)) # Sort the files by name in descending order to mimic 'reverse=True' request_files.sort(reverse=True) # Select the correct request file based on 'status' and 'precision' request_file = None for request_file in request_files: with request_file.open("r") as f: req_content = json.load(f) if req_content["status"] == "FINISHED" and req_content["precision"] == precision.split(".")[-1]: request_file = str(request_file) # Return empty string if no file found that matches criteria return request_file def get_raw_eval_results(results_path: str, requests_path: str, dynamic_path: str) -> list[EvalResult]: """From the path of the results folder root, extract all needed info for results""" with open(dynamic_path) as f: dynamic_data = json.load(f) results_path = Path(results_path) model_files = list(results_path.rglob('results_*.json')) model_files.sort(key=lambda file: parse_datetime(file.stem.removeprefix("results_"))) eval_results = {} # Wrap model_files iteration with tqdm for progress display for model_result_filepath in tqdm(model_files, desc="Processing model files"): # Creation of result eval_result = EvalResult.init_from_json_file(model_result_filepath) with logging_redirect_tqdm(): eval_result.update_with_request_file(requests_path) if eval_result.full_model in dynamic_data: eval_result.update_with_dynamic_file_dict(dynamic_data[eval_result.full_model]) # Hardcoding because of gating problem if any([org in eval_result.full_model for org in ["meta-llama/", "google/", "tiiuae/"]]): eval_result.still_on_hub = True # Store results of same eval together eval_name = eval_result.eval_name if eval_name in eval_results.keys(): eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None}) else: eval_results[eval_name] = eval_result results = [] for k, v in eval_results.items(): try: if v.status == "FINISHED": v.to_dict() # we test if the dict version is complete results.append(v) except KeyError as e: logging.error(f"Error while checking model {k} {v.date} json, no key: {e}") # not all eval values present continue return results