Emulated-Inc/general-knowledge-mcq-training-pool

General knowledge multiple-choice training pool

Public multiple-choice questions in medicine and health, law, history, philosophy, business and everyday general knowledge, from four datasets, 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, 236665 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.

Worth knowing. No row here offers more than five options, so the answer letter never runs past E: 479 rows carry three options, 175757 rows carry four options, 60429 rows carry five options. A recipe that has to read a longer option list, or write a letter beyond E, needs material this pool does not contain.

sources/

The same data untouched, 255198 rows, one directory per source, holding the files at the paths, in the 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 a CommonsenseQA question was built around, 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. Subjects: health. From openlifescienceai/medmcqa at revision 91c6572c454088bf71b679ad90aa8dffcd0d5868, files data/train-00000-of-00001.parquet, data/validation-00000-of-00001.parquet. 184723 rows here, of which 166414 also appear in pool.jsonl. 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. Its subject labels are medical specialities, so the subject field selects within health rather than across the eight subjects.

sources/lex_glue_case_hold

The CaseHOLD task of LexGLUE: a passage from a United States court opinion with the holding of the case it cites blanked out, and five candidate holdings of which one is the real one, taken from the Harvard Case Law corpus. Subjects: law. From coastalcph/lex_glue at revision c23fdff1a6bf74e0e1a71cb86f1e781d37da888c, files case_hold/train-00000-of-00001.parquet, case_hold/validation-00000-of-00001.parquet. 48880 rows here, of which 48842 also appear in pool.jsonl. Provenance class human, licence cc-by-4.0. Its own fields are context, endings as a list of five candidate holdings, label as the index of the correct one.

Worth knowing. The options are judicial holdings rather than short answers, so the rows are long, and the context is a citing passage rather than a question sentence. It is the hardest law material in the pool and the closest in shape to a professional law examination question.

sources/exams

High-school examination questions collected from the official state examinations of several ministries of education, over 24 subjects. Subjects: history, philosophy, business, geography and others. From mhardalov/exams at revision 4ff10804abb3341f8815cacd778181177bba7edd, files multilingual/train-00000-of-00001.parquet, multilingual/validation-00000-of-00001.parquet. 10633 rows here, of which 10495 also appear in pool.jsonl. Provenance class human, licence cc-by-sa-4.0. Its own fields are id, question with a stem and choices holding parallel text, label and para lists, answerKey, and info with the grade, the subject and the language.

Worth knowing. The multilingual configuration only, and its own test split is left out. No row is in English, and the subject field is how to select history, philosophy or business from the sciences it also carries.

sources/commonsense_qa

Crowdworker-written questions about everyday concepts, each with five options drawn from the neighbours of one ConceptNet concept so that the distractors are related rather than arbitrary. Subjects: general knowledge. From tau/commonsense_qa at revision 94630fe30dad47192a8546eb75f094926d47e155, files data/train-00000-of-00001.parquet, data/validation-00000-of-00001.parquet. 10962 rows here, of which 10914 also appear in pool.jsonl. Provenance class human, licence mit. Its own fields are id, question, question_concept, choices with parallel label and text lists, answerKey.

Worth knowing. Everyday reasoning rather than examined knowledge, which is the closest public material to the general-knowledge subject and is easier than it. Its published test split carries no answer key and is left out.

Provenance and licences

Every row was written by a person. All four sources are questions written for human examinations, written by judges in the course of deciding cases, or written by crowdworkers, which is the provenance class human. Nothing in the pool was generated by a model.

The pool as a whole is offered under cc-by-sa-4.0, which is the most restrictive term its sources compose to. The sources themselves are apache-2.0 for MedMCQA, cc-by-4.0 for CaseHOLD, cc- by-sa-4.0 for EXAMS and mit for CommonsenseQA. None of them is NonCommercial and none carries an unknown licence tag. Each rewritten row carries its own in the licence field and each directory under sources/ is one source, so a subset under a single licence can be selected. Attribution for the share-alike source goes to the authors of EXAMS, the CaseHOLD passages come from the Harvard Case Law corpus of United States court opinions, and questions taken from national and professional examinations remain the property of the boards that wrote them.

Filtering

Each source contributes its train split and its validation split, and nothing else. Within those splits rows were removed before publication, from both layers alike, by a word 8-gram overlap check (2160 rows), a whole-question match (10 rows) and an embedding similarity check (132 rows). Nothing else was filtered: no subject, no language and no difficulty was selected for or against, so the pool still holds the science questions its multi-subject sources carry alongside the humanities, and the subject field is how to select what you want.