Emulated-Inc/medical-mcq-training-pool

Medical multiple-choice training pool

Public multiple-choice medical questions from three repositories, read at the pinned revisions named below and laid out twice. Train on either layer or on both.

pool.jsonl

Every source rewritten into one shape, 219695 rows, one JSON object per line, with these fields.

The options of every row were permuted once when this file was built, under the recorded seed 20260912, and the answer letter is written in the permuted order, so that a source which publishes its correct option in a fixed position cannot teach that position. Rows are deduplicated across sources on the question together with its option set, keeping the first source that carries the question in the order of the sections below.

Two rules qualify that. The first is that a row one of whose options points at the others rather than answering the question keeps the option order its source published and the gold letter that goes with it, because permuting it would leave the pointer aimed at whatever landed above it. An option counts as a pointer when it is a bare All, None, Both, Neither or Either, with or without a full stop; or a phrase starting with one of those or with Any that reaches the above, these, those, them, the options, the choices, the answers or the statements, with the common misspellings of the connecting words and of above allowed for, and with a bare numeral counting as a quantifier only where it counts the other options as in two of the above; or a pointer written as option letters, A+B or a, b, g or both a and c. The phrase has to start within the first three words of the option, which is what keeps an ordinary answer such as two standard deviation above the mean out of the rule. 12843 rows here are in their source's order for that reason and the rest are permuted. Those rows keep their source's positional lean along with its order: their gold letters are A 3404, B 2738, C 2423, D 4270, E 8, against a flat distribution over the permuted rows.

The second is that a row two of whose options say the same thing once whitespace and case are set aside is not in this file at all, because its key names one letter while another letter is equally correct: 774 rows were dropped for that. Those rows are still in the untouched layer, which restates no answer key of its own.

sources/

The same data untouched, 238859 rows, one directory per source, holding the files at the paths, in the file format and with the columns its own repository publishes. Nothing here was renamed, reshaped, reordered or deduplicated. Use this layer if you want a field the rewritten one drops, such as the written explanations in MedMCQA or the concept phrases in MedQA, or if you would rather order the options yourself.

The sources

sources/medmcqa

Indian medical entrance examination questions, AIIMS and NEET PG, 1991 to the present, with the examining body's own answer key. From openlifescienceai/medmcqa at revision 91c6572c454088bf71b679ad90aa8dffcd0d5868, files data/train-00000-of-00001.parquet, data/validation-00000-of-00001.parquet. 186399 rows here, of which 167611 also appear in pool.jsonl. Language English, provenance class human, licence apache-2.0. Its own fields are id, question, opa to opd for the four options, cop for the zero-based index of the correct one, choicetype, exp for a written explanation, subjectname, topic_name.

Worth knowing. The apache-2.0 tag covers the packaging rather than the examination boards' copyright in the underlying questions, which the dataset card does not address. The published test split carries no answer key and is not here.

sources/medqa_usmle

United States medical licensing examination questions in their four-option English form, most of them clinical vignettes of several hundred words. From bigbio/med_qa at revision 484a6c066fe8e75c83edea0c88b5169316714fcd, files med_qa_en_4options_source/train-00000-of-00001.parquet. 9746 rows here, of which 9744 also appear in pool.jsonl. Language English, provenance class human, licence mit. Its own fields are metainfo naming the examination step, question, options as key and value records, answeridx naming the correct key, answer repeating its text, metamap_phrases.

Worth knowing. The Hub repository these files are read from states no licence, so the licence recorded on every row is the one the corpus's own repository publishes it under. No licence speaks to the examining boards' copyright in the questions themselves. Only the train split is included.

sources/medqa_mainland

Mainland China national medical licensing examination questions in their four-option form, in simplified Chinese. From bigbio/med_qa at revision 484a6c066fe8e75c83edea0c88b5169316714fcd, files med_qa_zh_4options_source/train-00000-of-00001.parquet. 27400 rows here, of which 27048 also appear in pool.jsonl. Language Chinese, provenance class human, licence mit. Its own fields are metainfo naming the textbook chapter, question, options as key and value records, answeridx naming the correct key, answer repeating its text.

Worth knowing. The Hub repository these files are read from states no licence, so the licence recorded on every row is the one the corpus's own repository publishes it under. No licence speaks to the examining boards' copyright in the questions themselves. Only the train split is included.

sources/medqa_taiwan

Taiwan medical licensing examination questions, in traditional Chinese. From bigbio/med_qa at revision 484a6c066fe8e75c83edea0c88b5169316714fcd, files med_qa_tw_source/train-00000-of-00001.parquet. 11298 rows here, of which 11289 also appear in pool.jsonl. Language Chinese, provenance class human, licence mit. Its own fields are metainfo naming the examination, question, options as key and value records, answeridx naming the correct key, answer repeating its text.

Worth knowing. The Hub repository these files are read from states no licence, so the licence recorded on every row is the one the corpus's own repository publishes it under. No licence speaks to the examining boards' copyright in the questions themselves. Only the train split is included.

sources/headqa

Spanish healthcare specialisation examinations, the state examinations sat by medicine, nursing, pharmacology, psychology, biology and chemistry graduates, in their English translation. From openlifescienceai/headqa at revision 38ce88d993b9e1c018de65169812f64e6f3289ef, files train.json, validation.json. 4016 rows here, of which 4003 also appear in pool.jsonl. Language English, provenance class human, licence mit. Its own fields are id, topic_name, and a data record holding Question, Options as a letter to text map, Correct Answer as the text and Correct Option as the letter.

Worth knowing. The re-upload these files are read from states no licence. The authors publish the corpus under MIT where they distribute it themselves, so redistribution here rests on that rather than on a statement attached to these files. Its subjects run past medicine into biology, chemistry and psychology, and the topic field is how to select. The published test split is not here.

Provenance and licences

Every row was written by a person. All three repositories publish questions written for human medical examinations, sat by candidates for a licence or a place, which is the provenance class human. Nothing in the pool was generated by a model.

Every source here is permissive and every row carries the licence it is under with a citation to the file that states it: apache-2.0 for MedMCQA from the tag on its own repository, MIT for MedQA from the LICENSE of github.com/jind11/MedQA, MIT for HeadQA from the LICENSE of github.com/aghie/head-qa. The two Hub re-uploads state no licence of their own, which is why the citation is to the corpus's own repository, and a source whose licence could not be cited that way is not here. None is share-alike and none is non-commercial, so the pool composes to no restrictive term. It is offered as other all the same, because no licence in this chain speaks to the examining boards' copyright in the underlying questions.

Filtering

Rows whose question duplicated or closely paraphrased a question in a held-out evaluation set were removed before publication, from both layers alike, by a word 8-gram overlap check (938 rows) followed by an embedding similarity check (107 rows). That evaluation set is not distributed here. Nothing else was filtered: no subject, no language and no difficulty was selected for or against, so the pool still holds the biology, chemistry and psychology questions its multi-subject sources carry alongside the clinical ones, and the subject field is how to select what you want.