CinderellaMan/rstar-coder-verified-io-deduped

Verified Python stdin/stdout dataset

A curated, deterministically verifiable subset of microsoft/rStar-Coder.

Repository: CinderellaMan/rstar-coder-verified-io-deduped

License

CC BY 4.0, inherited from the upstream dataset (microsoft/rStar-Coder, CC BY 4.0). You must give appropriate credit and indicate changes.

Attribution & changes

Derived from microsoft/rStar-Coder. Changes from upstream:

• filtered to a subset whose solution deterministically reproduces its testcases' outputs (5 runs, byte-identical after normalization);
• outputs re-derived and normalized (the shipped expected is the solution's normalized stdout, not the raw stored value);
• near-duplicate tasks merged — to fold producer-multiplied task variations (the same problem, lightly tweaked) into one task, via four signals: shared testcases with agreeing outputs; identical canonical source (AST-normalized, ignoring names/formatting); source similarity (Jaro >= 0.8 on long candidate pairs, where candidates come from both seed-adjacency and MinHash/LSH content similarity, so distant-seed variants are caught); and behavioral equivalence (each solution reproduces all the other's testcases — catches variations with a genuinely different solution and vetoes short boilerplate-similar matches); see provenance.json (post_selection_dedup);
• rows with fewer than 6 testcases dropped (v1): every shipped task has >= 6 verified testcases;
• problem statements removed — only solution source + I/O testcases are shipped.

Schema (one row per task)

id, source, solution_sha256, testcases (JSON: stdin byte-exact, expected normalized canonical, n_stored_outputs), num_testcases, meta (JSON: member_qids, alternate_solution_shas, source_repo, source_revision).

Interpreter

Verified under Python 3.14.5. Outputs are interpreter-version sensitive; compare with the same conservative normalizer (see provenance.json).