import torch import os import json import gc import csv from PIL import Image from datasets import load_dataset,load_from_disk from collections import defaultdict from tqdm import tqdm from ..utils import save_json,extract,judge_multi_choice from ..base_dataset import BaseDataset from ..question_formats import get_multiple_choice_prompt medical_subject = ["anatomy","clinical_knowledge","college_medicine","genetics","traditional_chinese_medicine","nutrition","virology"] class CMMLU(BaseDataset): def __init__(self,model,dataset_path,output_path): self.model = model self.output_path = output_path self.dataset_path = dataset_path if dataset_path else "haonan-li/cmmlu" self.samples = [] self.chunk_idx = int(os.environ.get("chunk_idx",0)) self.num_chunks = int(os.environ.get("num_chunks",1)) def load_data(self): dataset = load_dataset(self.dataset_path)["train"] for idx,sample in tqdm(enumerate(dataset)): if idx % self.num_chunks == self.chunk_idx: if sample["task"] in medical_subject: sample = self.construct_messages(sample) self.samples.append(sample) print(len(self.samples)) return self.samples def construct_messages(self,sample): answer = sample["answer"] OptionA = sample["A"] OptionB = sample["B"] OptionC = sample["C"] OptionD = sample["D"] question = sample["question"] choices = [OptionA,OptionB,OptionC,OptionD] choices = [f"{chr(65+i)}.{choices[i]}" for i in range(len(choices))] is_reasoning = True if os.environ.get("REASONING","False") == "True" else False prompt = get_multiple_choice_prompt(question,choices,is_reasoning,lang = "zh") messages = {"prompt":prompt} sample["prompt"] = prompt sample["messages"] = messages sample["answer"] = answer sample["choices"] = choices # import pdb;pdb.set_trace() return sample def cal_metrics(self,out_samples): total = 0 right = 0 for i,sample in enumerate(out_samples): response = sample["response"] response = extract(response,"answer") choices = sample["choices"] answer = sample["answer"] correct = judge_multi_choice(choices,answer,response) out_samples[i]["correct"] = correct if correct: right += 1 total += 1 metrics = {"total metrics":{"total":total,"right":right,"acc":right/total}} return metrics,out_samples