competitive-code-curated

Derived from:

  DeepCoder-Preview-Dataset, Agentica
  https://huggingface.co/datasets/agentica-org/DeepCoder-Preview-Dataset
  revision 177913a7bd43791646ef6a43645caa3c871ab3db, configs `taco` and
  `primeintellect` (train)
  Licensed under the MIT licence.

whose rows come from TACO (BAAI, https://huggingface.co/datasets/BAAI/TACO,
Apache-2.0) and PrimeIntellect's verifiable coding problems
(https://huggingface.co/datasets/PrimeIntellect/verifiable-coding-problems).
Problem statements were originally published by their contest judges
(Codeforces, AtCoder, AIZU, CodeChef, HackerEarth and others) and stay
subject to those judges' terms.

Changes made: stdin/stdout problems only (function-call problems removed);
problems whose statement admits several answers ("print any", several valid
answers without a tie-break) removed; problems whose tests contradict an
example of their own statement removed (every copy); problems overlapping
LiveCodeBench (livecodebench/code_generation_lite @0fe84c39) removed;
duplicates and near-duplicates merged; each problem kept only if one of its
first three Python 3 upstream solutions passes all its tests and none of them
disagrees; tests subsampled under a time budget and a per-problem time limit
derived from the reference's CPU time; problems split into train, eval and
qualification by problem id. One passing upstream solution per problem is
shipped in references.parquet.

Every row carries its upstream source and row id in `origin_json`. This
attribution must be kept with the data.