math-intuition-20260908-402-easy-10

A 10-seed downsample of amphora/math-intuition-20260906-403-easy-30.

• 4,020 rows — 402 families x 10 seeds, at the easy preset
• Seeds kept are indices ::00-::09 of each family, so the subset is deterministic and reproducible
• Family 1309.6504 is excluded: its rendered questions are 235,000 characters (142k tokens), past the context window of every model under evaluation, so it could never be scored

Each question is self-contained and states its own answer contract, ending with an <answer>{...}</answer> block plus a worked example. Answers are graded by the originating family's own parse_answer() / verify() rather than by string match.

instances/

instances/math-intuition-20260908-402-easy-10.train.pkl holds the pickled generator instance for every row, so a grader can call verify() without re-running the generators.

Instances are built from each row's params column, not from DIFFICULTY[preset] — for some families the preset rung in the source repo no longer matches the kwargs the row was generated with, and params records the truth. All 4,020 instances were re-rendered and compared byte-for-byte against the published question before upload.

The pickle stores ids as a plain list. Storing a lazy datasets Column instead embeds an absolute path to the Arrow file on the machine that built it, which makes the cache unopenable anywhere else.

Known issue

search_space contains integers up to ~2,500 digits, which overflows int64, so load_dataset() fails on the CSV. Read it with pandas.read_csv(..., dtype=str) or supply an explicit string feature spec for that column.