import glob import json import os from dataclasses import dataclass import dateutil import numpy as np from src.display.formatting import make_clickable_model from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType from src.submission.check_validity import is_model_on_hub @dataclass class EvalResult: """Represents one full evaluation. Built from a combination of the result and request file for a given run.""" eval_name: str # org_model_precision (uid) full_model: str # org/model (path on hub) org: str model: str revision: str # commit hash, "" if main results: dict precision: Precision = Precision.Unknown model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ... weight_type: WeightType = WeightType.Original # Original or Adapter architecture: str = "Unknown" license: str = "?" likes: int = 0 num_params: int = 0 date: str = "" # submission date of request file still_on_hub: bool = False energy_score: str = "NA" # energy consumption in kWh, "NA" if not available @classmethod def init_from_json_file(self, json_filepath): """Inits the result from the specific model result file""" with open(json_filepath) as fp: data = json.load(fp) config = data.get("config_general") # Handle case where config_general is None if config is None: # Set default values precision = Precision.Unknown org_and_model = data.get("model_name", "Unknown/Unknown") if not isinstance(org_and_model, str): org_and_model = "Unknown/Unknown" revision = "main" else: # Precision precision = Precision.from_str(config.get("model_dtype")) # Get revision revision = config.get("model_sha", "") # Get model and org org_and_model = config.get("model_name", config.get("model_args", None)) if isinstance(org_and_model, str): org_and_model = org_and_model.split("/", 1) else: org_and_model = ["Unknown", "Unknown"] # Already handled above 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) # Use a safe default for model_sha if config is None model_sha = "main" if config is not None: model_sha = config.get("model_sha", "main") still_on_hub, _, model_config = is_model_on_hub( full_model, model_sha, trust_remote_code=True, test_tokenizer=False ) architecture = "?" if model_config is not None: architectures = getattr(model_config, "architectures", None) if architectures: architecture = ";".join(architectures) # Extract results available in this file (some results are split in several files) results = {} # Check if results key exists in the data if "results" not in data: # If no results, set all benchmarks to None for task in Tasks: task = task.value results[task.benchmark] = None else: # Process results normally for task in Tasks: task = task.value # We average all scores of a given metric (not all metrics are present in all files) # Handle metrics that could be None in the JSON metric_values = [] # Define the expected metric name and alternative names for each benchmark expected_metric = task.metric alternative_metrics = [] print(f"Processing benchmark: {task.benchmark}, expected metric: {expected_metric}") # Set up alternative metric names based on the benchmark if task.benchmark == "custom|folio:logical_reasoning|0": if expected_metric != "folio_em": alternative_metrics = ["folio_em"] elif task.benchmark == "custom|telecom:qna|0": if expected_metric != "telecom_qna_em": alternative_metrics = ["telecom_qna_em"] elif task.benchmark == "custom|3gpp:tsg|0": if expected_metric != "em": alternative_metrics = ["em"] elif task.benchmark == "custom|math:problem_solving|0": if expected_metric != "math_metric": alternative_metrics = ["math_metric"] elif task.benchmark == "custom|spider:text2sql|0": if expected_metric != "sql_metric": alternative_metrics = ["sql_metric"] # Check for results with the benchmark name for k, v in data["results"].items(): if task.benchmark == k: # Try the expected metric name first metric_value = v.get(expected_metric) # If not found, try alternative metric names if metric_value is None: for alt_metric in alternative_metrics: if alt_metric in v: metric_value = v.get(alt_metric) break if metric_value is not None: metric_values.append(metric_value) print(f"Found metric value for {task.benchmark}: {metric_value}") accs = np.array([v for v in metric_values if v is not None]) if len(accs) == 0: # Also check the "all" section for metrics if "all" in data["results"]: all_results = data["results"]["all"] print(f"Checking 'all' section for {task.benchmark}, available keys: {list(all_results.keys())}") # Try the expected metric name first metric_value = all_results.get(expected_metric) # If not found, try alternative metric names if metric_value is None: for alt_metric in alternative_metrics: if alt_metric in all_results: metric_value = all_results.get(alt_metric) print(f"Found alternative metric {alt_metric} in 'all' section") break if metric_value is not None: accs = np.array([metric_value]) print(f"Found metric value in 'all' section for {task.benchmark}: {metric_value}") else: results[task.benchmark] = None continue else: results[task.benchmark] = None continue mean_acc = np.mean(accs) * 100.0 results[task.benchmark] = mean_acc print(f"Final result for {task.benchmark}: {mean_acc}") # Extract energy score if available energy_score = "NA" if "energy_metrics" in data and data["energy_metrics"] is not None and data["energy_metrics"].get("enabled", False): total_energy = data["energy_metrics"].get("total_energy", 0) if total_energy > 0: energy_score = f"{total_energy:.5f}" return self( eval_name=result_key, full_model=full_model, org=org, model=model, results=results, precision=precision, revision=revision, still_on_hub=still_on_hub, architecture=architecture, energy_score=energy_score ) def update_with_request_file(self, requests_path): """Finds the relevant request file for the current model and updates info with it""" request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name) try: with open(request_file, "r") as f: request = json.load(f) self.model_type = ModelType.from_str(request.get("model_type", "")) self.weight_type = WeightType[request.get("weight_type", "Original")] self.license = request.get("license", "?") self.likes = request.get("likes", 0) self.num_params = request.get("params", 0) self.date = request.get("submitted_time", "") self.architecture = request.get("architectures", "Unknown") # delete later self.status = request.get("status", "FAILED") except Exception: self.status = "FAILED" print(f'Could not find request file for {self.org}/{self.model} with "precision:{self.precision.value.name},model_type:{self.model_type}",license:{self.license},status:{self.status}') def to_dict(self): """Converts the Eval Result to a dict compatible with our dataframe display""" available_metrics = [v for v in self.results.values() if v is not None] average = sum(available_metrics) / len([v for v in available_metrics if v is not None]) if available_metrics else None 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.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.energy_score.name: self.energy_score, } print(f"\nConverting to dict for model: {self.full_model}") for task in Tasks: result = self.results.get(task.value.benchmark) print(f" Task: {task.value.col_name}, Benchmark: {task.value.benchmark}, Result: {result}") data_dict[task.value.col_name] = "NA" if result is None else round(result, 2) 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""" request_files = os.path.join( requests_path, f"{model_name}_eval_request_*.json", ) request_files = glob.glob(request_files) # Select correct request file (precision) request_file = "" request_files = sorted(request_files, reverse=True) for tmp_request_file in request_files: with open(tmp_request_file, "r") as f: req_content = json.load(f) if ( req_content["status"] in ["FINISHED"] and req_content["precision"] == precision.split(".")[-1] ): request_file = tmp_request_file return request_file def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]: """From the path of the results folder root, extract all needed info for results""" model_result_filepaths = [] for root, _, files in os.walk(results_path): # We should only have json files in model results if len(files) == 0 or any([not f.endswith(".json") for f in files]): continue # Sort the files by date try: files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7]) except dateutil.parser._parser.ParserError: files = [files[-1]] for file in files: model_result_filepaths.append(os.path.join(root, file)) eval_results = {} for model_result_filepath in model_result_filepaths: try: # Creation of result print(f"\nProcessing file: {model_result_filepath}") eval_result = EvalResult.init_from_json_file(model_result_filepath) # Skip entries with Unknown/Unknown model name if eval_result.full_model == "Unknown/Unknown": print(f"Skipping invalid result file: {model_result_filepath}") continue print(f"Model: {eval_result.full_model}") print(f"Results before update_with_request_file:") for benchmark, value in eval_result.results.items(): print(f" {benchmark}: {value}") eval_result.update_with_request_file(requests_path) print(f"Results after update_with_request_file:") for benchmark, value in eval_result.results.items(): print(f" {benchmark}: {value}") # Store results of same eval together eval_name = eval_result.eval_name if eval_name in eval_results.keys(): print(f"Updating existing results for {eval_name}") eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None}) else: print(f"Adding new results for {eval_name}") eval_results[eval_name] = eval_result except Exception as e: print(f"Error processing result file {model_result_filepath}: {str(e)}") continue results = [] for v in eval_results.values(): try: v.to_dict() # we test if the dict version is complete results.append(v) except KeyError: # not all eval values present continue return results