Datasets:

horizon-llm
/

USACO-Judge

license: cc0-1.0
language:
  - en
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - code
  - competitive-programming
  - verification
  - usaco
configs:
  - config_name: problems
    data_files: problems.jsonl
  - config_name: submissions
    data_files: submissions.jsonl

USACO-Judge

Paper: Beyond the Verdict: Agentic Code Judging with Verifiable Feedback
Code: https://github.com/horizon-llm/IOI2026
Project page: https://horizon-llm.github.io/IOI2026/

A benchmark for judging competitive-programming solutions: given a problem and a candidate solution, decide Accept / Reject, and on Reject, produce a concrete input that breaks the code. Existing hacking benchmarks are built entirely from known-wrong candidates, so they only test the breaking half of verification. USACO-Judge is balanced 1:2 AC:non-AC, so it also tests whether a verifier correctly accepts solutions that are actually correct. Built from 39 official USACO 2026 problems and 900 real candidate submissions.

`problems` (39 rows)

One row per problem: cpid (USACO's official problem id — usaco.org/index.php?page=viewproblem2&cpid=<cpid>), title, contest, division, problem_number, source_url, statement_md, checker_kind, time_limit_s, memory_limit_mb, n_tests, n_sample_tests, point_scores, val_cpp (validator), std_cpp (reference solution), chk_cpp (custom checker, null for problems that use a standard comparator), std_source_submission_id, std_origin.

`submissions` (900 rows: 270 easy + 630 hard)

One row per submission: sid, cpid (links to problems), split (easy/hard), code, verdict (AC/WA/TLE/MLE/RE), score (0-100). 300 are AC, 600 are not. The easy split draws from the 18 Bronze/Silver problems, hard from the 21 Gold/Platinum/Open problems.

Non-AC candidates are stratified across five score bands (1-19, 20-39, 40-59, 60-79, 80-99), each a distinct plateau of suboptimal algorithms rather than a single bug pattern.

License

Released under CC0 1.0, matching the code repository. Problem statements come from the official USACO 2026 contests (see source_url).

Citation

@misc{deng2026beyond,
  title         = {Beyond the Verdict: Agentic Code Judging with Verifiable Feedback},
  author        = {Deng, Zheye and Li, Sha and Yu, Changlong and Liu, Xin and Lu, Qin and Nigam, Priyanka and Song, Yangqiu},
  year          = {2026},
  eprint        = {2610.18760},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.18760}
}

USACO-Judge

problems (39 rows)

submissions (900 rows: 270 easy + 630 hard)

License

Citation

`problems` (39 rows)

`submissions` (900 rows: 270 easy + 630 hard)