CPunisher/JavaBench

1import glob2import json3import math4import os5from dataclasses import dataclass6 7import dateutil8import numpy as np9 10from src.display.formatting import make_clickable_model11from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType12from src.submission.check_validity import is_model_on_hub13 14 15@dataclass16class EvalResult:17    """Represents one full evaluation. Built from a combination of the result and request file for a given run.18    """19    eval_name: str # org_model_precision (uid)20    full_model: str # org/model (path on hub)21    org: str 22    model: str23    revision: str # commit hash, "" if main24    results: dict25    precision: Precision = Precision.Unknown26    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...27    weight_type: WeightType = WeightType.Original # Original or Adapter28    architecture: str = "Unknown" 29    license: str = "?"30    likes: int = 031    num_params: int = 032    date: str = "" # submission date of request file33    still_on_hub: bool = False34 35    @classmethod36    def init_from_json_file(self, json_filepath):37        """Inits the result from the specific model result file"""38        with open(json_filepath) as fp:39            data = json.load(fp)40 41        config = data.get("config")42 43        # Precision44        precision = Precision.from_str(config.get("model_dtype"))45 46        # Get model and org47        org_and_model = config.get("model_name", config.get("model_args", None))48        org_and_model = org_and_model.split("/", 1)49 50        if len(org_and_model) == 1:51            org = None52            model = org_and_model[0]53            result_key = f"{model}_{precision.value.name}"54        else:55            org = org_and_model[0]56            model = org_and_model[1]57            result_key = f"{org}_{model}_{precision.value.name}"58        full_model = "/".join(org_and_model)59 60        still_on_hub, _, model_config = is_model_on_hub(61            full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False62        )63        architecture = "?"64        if model_config is not None:65            architectures = getattr(model_config, "architectures", None)66            if architectures:67                architecture = ";".join(architectures)68 69        # Extract results available in this file (some results are split in several files)70        results = {}71        for task in Tasks:72            task = task.value73 74            # We average all scores of a given metric (not all metrics are present in all files)75            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])76            if accs.size == 0 or any([acc is None for acc in accs]):77                continue78 79            mean_acc = np.mean(accs) * 100.080            results[task.benchmark] = mean_acc81 82        return self(83            eval_name=result_key,84            full_model=full_model,85            org=org,86            model=model,87            results=results,88            precision=precision,  89            revision= config.get("model_sha", ""),90            still_on_hub=still_on_hub,91            architecture=architecture92        )93 94    def update_with_request_file(self, requests_path):95        """Finds the relevant request file for the current model and updates info with it"""96        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)97 98        try:99            with open(request_file, "r") as f:100                request = json.load(f)101            self.model_type = ModelType.from_str(request.get("model_type", ""))102            self.weight_type = WeightType[request.get("weight_type", "Original")]103            self.license = request.get("license", "?")104            self.likes = request.get("likes", 0)105            self.num_params = request.get("params", 0)106            self.date = request.get("submitted_time", "")107        except Exception:108            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")109 110    def to_dict(self):111        """Converts the Eval Result to a dict compatible with our dataframe display"""112        average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)113        data_dict = {114            "eval_name": self.eval_name,  # not a column, just a save name,115            AutoEvalColumn.precision.name: self.precision.value.name,116            AutoEvalColumn.model_type.name: self.model_type.value.name,117            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,118            AutoEvalColumn.weight_type.name: self.weight_type.value.name,119            AutoEvalColumn.architecture.name: self.architecture,120            AutoEvalColumn.model.name: make_clickable_model(self.full_model),121            AutoEvalColumn.revision.name: self.revision,122            AutoEvalColumn.average.name: average,123            AutoEvalColumn.license.name: self.license,124            AutoEvalColumn.likes.name: self.likes,125            AutoEvalColumn.params.name: self.num_params,126            AutoEvalColumn.still_on_hub.name: self.still_on_hub,127        }128 129        for task in Tasks:130            data_dict[task.value.col_name] = self.results[task.value.benchmark]131 132        return data_dict133 134 135def get_request_file_for_model(requests_path, model_name, precision):136    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""137    request_files = os.path.join(138        requests_path,139        f"{model_name}_eval_request_*.json",140    )141    request_files = glob.glob(request_files)142 143    # Select correct request file (precision)144    request_file = ""145    request_files = sorted(request_files, reverse=True)146    for tmp_request_file in request_files:147        with open(tmp_request_file, "r") as f:148            req_content = json.load(f)149            if (150                req_content["status"] in ["FINISHED"]151                and req_content["precision"] == precision.split(".")[-1]152            ):153                request_file = tmp_request_file154    return request_file155 156 157def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:158    """From the path of the results folder root, extract all needed info for results"""159    model_result_filepaths = []160 161    for root, _, files in os.walk(results_path):162        # We should only have json files in model results163        if len(files) == 0 or any([not f.endswith(".json") for f in files]):164            continue165 166        # Sort the files by date167        try:168            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])169        except dateutil.parser._parser.ParserError:170            files = [files[-1]]171 172        for file in files:173            model_result_filepaths.append(os.path.join(root, file))174 175    eval_results = {}176    for model_result_filepath in model_result_filepaths:177        # Creation of result178        eval_result = EvalResult.init_from_json_file(model_result_filepath)179        eval_result.update_with_request_file(requests_path)180 181        # Store results of same eval together182        eval_name = eval_result.eval_name183        if eval_name in eval_results.keys():184            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})185        else:186            eval_results[eval_name] = eval_result187 188    results = []189    for v in eval_results.values():190        try:191            v.to_dict() # we test if the dict version is complete192            results.append(v)193        except KeyError:  # not all eval values present194            continue195 196    return results197