Exclude all 15 source time_session rows that omit duration information, then sample 500 of the 985 eligible rows with random.Random(42) and restore source order. Removes all nine previously selected defective questions without synthesizing facts from gold answers. Original row IDs, source text, choice order and source labels preserved. Other six split Parquet files and every raw source unchanged byte-for-byte. All 500 selected answers covered by independent arithmetic checks; 16 pinned tokenizers and offline rebuild verified. Source-inherited distractor bias remains unchanged.
Normalize PythonIO Input/Output layout and answer whitespace
Use LF and Input: {input} followed by Output: in context, retaining one leading ASCII space per choice. Remove trailing answer whitespace from 45 incorrect choices across 44 questions; preserve code indentation and whitespace inside output values. All IDs, labels, sampled rows, choice order, raw sources, and other six split Parquet files unchanged. CPU-only verification across all 16 pinned tokenizers; no model evaluation.
Add tiny_openbookqa using the full 500-question main/test split
Source XiangPan/pythonio-mc test, pinned revision a85256a18bccaaa3d66d9bf86247436948b15353. Same seed-42 curation; native Program/Input/Output question and output-text choices, with A-D keys mapped to indices. Program text, quoting and trailing output whitespace preserved; no source programs executed. Existing five splits unchanged byte-for-byte. Total 3000 evaluation-only questions. README remains concise; source card has no declared license.
Version 2.1.0. Source AxiomicLabs/LogicMark train, pinned revision e357c145b63fb4a96813bdbb5b07a62fdb97d958. Same seed-42 curation; native context/options and answer_index. Existing four splits unchanged byte-for-byte. Total 2500 evaluation-only questions; no model evaluation executed.
Replace collection with four seeded 500-question, four-choice evaluation splits
