Datasets:

HaolinRPI
/

medmcqa

language:
  - en
license: apache-2.0
pretty_name: MedMCQA
size_categories:
  - 10K<n<100K
source_datasets:
  - openlifescienceai/medmcqa
task_categories:
  - question-answering
  - multiple-choice
tags:
  - medical
  - medmcqa
dataset_info:
  - config_name: sft
    features:
      - name: prompt
        dtype: string
      - name: answer
        dtype: string
    splits:
      - name: train
        num_examples: 20000
      - name: validation
        num_examples: 200
  - config_name: train
    splits:
      - name: train
        num_examples: 20000
  - config_name: validation
    splits:
      - name: validation
        num_examples: 200
  - config_name: mia_members
    splits:
      - name: train
        num_examples: 2000
  - config_name: mia_nonmembers
    splits:
      - name: train
        num_examples: 2000
  - config_name: filtered
    splits:
      - name: train
        num_examples: 84403
configs:
  - config_name: sft
    data_files:
      - split: train
        path: sft/sft_train_20000.jsonl
      - split: validation
        path: sft/sft_validation_200.jsonl
  - config_name: train
    data_files:
      - split: train
        path: train_20000.jsonl
  - config_name: validation
    data_files:
      - split: validation
        path: validation_200.jsonl
  - config_name: mia_members
    data_files:
      - split: train
        path: mia_members_2000.jsonl
  - config_name: mia_nonmembers
    data_files:
      - split: train
        path: mia_nonmembers_2000.jsonl
  - config_name: filtered
    data_files:
      - split: train
        path: medmcqa_filtered.parquet

MedMCQA

由 openlifescienceai/medmcqa 的 train split 清洗、去重后划分得到。上游许可证为 Apache-2.0。

默认配置 sft 可直接用于监督微调。answer 是完整解答加正确选项字母，不是单独的选项标签。

from datasets import load_dataset

sft = load_dataset("HaolinRPI/medmcqa", "sft")
train = load_dataset("HaolinRPI/medmcqa", "train", split="train")
validation = load_dataset("HaolinRPI/medmcqa", "validation", split="validation")
members = load_dataset("HaolinRPI/medmcqa", "mia_members", split="train")
nonmembers = load_dataset("HaolinRPI/medmcqa", "mia_nonmembers", split="train")
filtered = load_dataset("HaolinRPI/medmcqa", "filtered", split="train")

划分

Training、MIA Non-member 和 Validation 的规范化 Question 互不重叠。MIA Member 是 Training 的子集。随机种子是 42。完整统计见 split_manifest.json。

文件	条数	用途
`sft/sft_train_20000.jsonl`	20,000	SFT 训练集，只含 `prompt` 和 `answer`
`sft/sft_validation_200.jsonl`	200	SFT 验证集，只含 `prompt` 和 `answer`
`train_20000.jsonl`	20,000	训练样本，保留原始字段
`mia_members_2000.jsonl`	2,000	Training 内部的 MIA Member
`mia_nonmembers_2000.jsonl`	2,000	不参与训练的 MIA Non-member
`validation_200.jsonl`	200	独立验证样本
`medmcqa_filtered.parquet`	84,403	全部合格记录，含未使用样本

SFT 文本

prompt 由题干、A–D 选项和指令 “Answer with the letter of the correct option.” 组成。answer 由解释和正确选项字母组成，字母单独占最后一行。