import glob import json import math import os from dataclasses import dataclass import dateutil import numpy as np from src.display.formatting import make_clickable_model # changes to be made here from src.display.utils import AutoEvalColumn, ModelType, ModelArch, Precision, HarnessTasks, WeightType, OpenEndedColumns, MedSafetyColumns, MedicalSummarizationColumns, ACIColumns, SOAPColumns 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 dataset_results: dict # changes to be made here open_ended_results: dict med_safety_results: dict medical_summarization_results: dict aci_results: dict soap_results: dict is_domain_specific: bool use_chat_template: bool # clinical_type_results:dict precision: Precision = Precision.Unknown model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ... weight_type: WeightType = WeightType.Original # Original or Adapter backbone:str = "Unknown" license: str = "?" likes: int = 0 num_params: int = 0 date: str = "" # submission date of request file still_on_hub: bool = False display_result:bool = True @classmethod def init_from_json_file(self, json_filepath, evaluation_metric): """Inits the result from the specific model result file""" with open(json_filepath) as fp: try: data = json.load(fp) except: breakpoint() config = data.get("config") # Precision precision = Precision.from_str(config.get("model_dtype")) model_type = ModelType.from_str(config.get("model_type", "")) license = config.get("license", "?") num_params = config.get("num_params", "?") display_result = config.get("display_result", True) display_result = False if display_result=="False" else True # Get model and org org_and_model = config.get("model_name", config.get("model_args", None)) org_and_model = org_and_model.split("/", 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) still_on_hub, _, model_config = is_model_on_hub( full_model, config.get("revision", "main"), trust_remote_code=True, test_tokenizer=False ) backbone = "?" if model_config is not None: backbones = getattr(model_config, "architectures", None) if backbones: backbone = ";".join(backbones) # Extract results available in this file (some results are split in several files) harness_results = {} if "closed-ended" in data["results"]: for task in HarnessTasks: task = task.value # We average all scores of a given metric (not all metrics are present in all files) try: accs = np.array([v.get(task.metric, None) for k, v in data["results"]["closed-ended"].items() if task.benchmark == k]) except: # breakpoint() accs = np.array([]) if accs.size == 0 or any([acc is None for acc in accs]): continue mean_acc = np.mean(accs) # * 100.0 harness_results[task.benchmark] = mean_acc open_ended_results = {} if "open-ended" in data["results"]: for task in OpenEndedColumns: task = task.value # We average all scores of a given metric (not all metrics are present in all files) accs = data["results"]["open-ended"]["overall"][task.benchmark] if task.benchmark in data["results"]["open-ended"]["overall"] else None open_ended_results[task.benchmark] = accs if open_ended_results["ELO_intervals"] is not None and open_ended_results["Score_intervals"] is not None: open_ended_results["ELO_intervals"] = "+" + str(open_ended_results["ELO_intervals"][1]) + "/-" + str(abs(open_ended_results["ELO_intervals"][0])) open_ended_results["Score_intervals"] = "+" + str(open_ended_results["Score_intervals"][1]) + "/-" + str(abs(open_ended_results["Score_intervals"][0])) # breakpoint() # changes to be made here med_safety_results = {} if "med-safety" in data["results"]: for task in MedSafetyColumns: task = task.value if task.benchmark == "Harmfulness Score": accs = data["results"]["med-safety"][task.benchmark] med_safety_results[task.benchmark] = accs elif task.benchmark == "95% CI": accs = data["results"]["med-safety"][task.benchmark] med_safety_results[task.benchmark] = "+" + str(round(accs[1], 3)) + "/-" + str(round(abs(accs[0]), 3)) else: accs = data["results"]["med-safety"][task.benchmark]["score"] med_safety_results[task.benchmark] = accs medical_summarization_results = {} if "medical-summarization" in data["results"]: for task in MedicalSummarizationColumns: task = task.value try: accs = np.array([v for k, v in data["results"]["medical-summarization"]["clinical_trial"].items() if task.benchmark == k]) except: accs = np.array([]) if accs.size == 0 or any([acc is None for acc in accs]): continue mean_acc = np.mean(accs) # * 100.0 medical_summarization_results[task.benchmark] = mean_acc aci_results = {} if "note-generation" in data["results"] and "aci" in data["results"]["note-generation"]: for task in ACIColumns: task = task.value try: accs = np.array([v for k, v in data["results"]["note-generation"]["aci"].items() if task.benchmark == k]) except: accs = np.array([]) if accs.size == 0 or any([acc is None for acc in accs]): continue mean_acc = np.mean(accs) # * 100.0 aci_results[task.benchmark] = mean_acc soap_results = {} if "note-generation" in data["results"] and "soap" in data["results"]["note-generation"]: for task in SOAPColumns: task = task.value try: accs = np.array([v for k, v in data["results"]["note-generation"]["soap"].items() if task.benchmark == k]) except: accs = np.array([]) if accs.size == 0 or any([acc is None for acc in accs]): continue mean_acc = np.mean(accs) # * 100.0 soap_results[task.benchmark] = mean_acc if open_ended_results == {} or med_safety_results == {} or medical_summarization_results == {} or aci_results == {} or soap_results == {}: open_ended_results = {} med_safety_results = {} medical_summarization_results = {} aci_results = {} soap_results = {} # types_results = {} # for clinical_type in ClinicalTypes: # clinical_type = clinical_type.value # # We average all scores of a given metric (not all metrics are present in all files) # accs = np.array([v.get(clinical_type.metric, None) for k, v in data[evaluation_metric]["clinical_type_results"].items() if clinical_type.benchmark == k]) # if accs.size == 0 or any([acc is None for acc in accs]): # continue # mean_acc = np.mean(accs) # * 100.0 # types_results[clinical_type.benchmark] = mean_acc return self( eval_name=result_key, full_model=full_model, org=org, model=model, revision=config.get("revision", ""), dataset_results=harness_results, open_ended_results=open_ended_results, med_safety_results=med_safety_results, medical_summarization_results=medical_summarization_results, aci_results=aci_results, soap_results=soap_results, is_domain_specific=config.get("is_domain_specific", False), # Assuming a default value use_chat_template=config.get("use_chat_template", False), # Assuming a default value precision=precision, model_type=model_type, weight_type=WeightType.from_str(config.get("weight_type", "")), # Assuming the default value backbone=backbone, license=license, likes=config.get("likes", 0), # Assuming a default value num_params=num_params, still_on_hub=still_on_hub, display_result=display_result, date=config.get("submitted_time","") ) 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.precision = request.get("precision", "float32") except Exception: pass # print( # f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}" # ) # print(f" Args used were - {request_file=}, {requests_path=}, {self.full_model=},") def to_dict(self, subset): """Converts the Eval Result to a dict compatible with our dataframe display""" 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 + (" 🏥" if self.is_domain_specific else ""), AutoEvalColumn.weight_type.name: self.weight_type.value.name, # AutoEvalColumn.architecture.name: self.architecture.value.name, # AutoEvalColumn.backbone.name: self.backbone, AutoEvalColumn.model.name: make_clickable_model(self.full_model), AutoEvalColumn.is_domain_specific.name: self.is_domain_specific, AutoEvalColumn.use_chat_template.name: self.use_chat_template, AutoEvalColumn.revision.name: self.revision, 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.date.name: self.date, "display_result" : self.display_result, } if subset == "datasets": average = sum([v for v in self.dataset_results.values() if v is not None]) / len(HarnessTasks) data_dict[AutoEvalColumn.average.name] = average if len(self.dataset_results) > 0: for task in HarnessTasks: data_dict[task.value.col_name] = self.dataset_results[task.value.benchmark] return data_dict if subset == "open_ended": if len(self.open_ended_results) > 0: for task in OpenEndedColumns: data_dict[task.value.col_name] = self.open_ended_results[task.value.benchmark] return data_dict # changes to be made here if subset == "med_safety": if len(self.med_safety_results) > 0: for task in MedSafetyColumns: data_dict[task.value.col_name] = self.med_safety_results[task.value.benchmark] return data_dict if subset == "medical_summarization": if len(self.medical_summarization_results) > 0: adjusted_conciseness = max(0, self.medical_summarization_results["brief"]) coverage = self.medical_summarization_results["coverage"] hm = 2 / (1/coverage + 1/adjusted_conciseness) if not (adjusted_conciseness == 0 or coverage == 0) else 0 conformity = self.medical_summarization_results["conform"] consistency = self.medical_summarization_results["fact"] overall = sum([hm, conformity, consistency]) / 3 data_dict[AutoEvalColumn.overall.name] = overall for task in MedicalSummarizationColumns: data_dict[task.value.col_name] = self.medical_summarization_results[task.value.benchmark] return data_dict if subset == "aci": overall = sum([v for v in self.aci_results.values() if v is not None]) / len(ACIColumns) data_dict[AutoEvalColumn.overall.name] = overall if len(self.aci_results) > 0: for task in ACIColumns: data_dict[task.value.col_name] = self.aci_results[task.value.benchmark] return data_dict if subset == "soap": overall = sum([v for v in self.soap_results.values() if v is not None]) / len(SOAPColumns) data_dict[AutoEvalColumn.overall.name] = overall if len(self.soap_results) > 0: for task in SOAPColumns: data_dict[task.value.col_name] = self.soap_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""" 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, evaluation_metric: 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: # Creation of result eval_result = EvalResult.init_from_json_file(model_result_filepath, evaluation_metric) # eval_result.update_with_request_file(requests_path) # 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 = [] # clinical_type_results = [] for v in eval_results.values(): try: v.to_dict(subset="dataset") # we test if the dict version is complete if not v.display_result: continue results.append(v) except KeyError: # not all eval values present continue # breakpoint() return results