Datasets:
| license: apache-2.0 | |
| task_categories: [text-generation] | |
| tags: [code, code-contests, zevo] | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: validation | |
| path: data/validation-*.parquet | |
| # zero-code L1 validation set — CodeContests valid (117) | |
| Independent validation for Zevo task `zero-code-l1` checkpoint selection. | |
| - Source: HuggingFace `deepmind/code_contests` split `valid` (117 problems). | |
| - Schema: same fields as the L1 LiveCodeBench test set (`id`, `question`, `public_test_cases`, `private_test_cases`, `metadata`). | |
| - Tests: CodeContests **public_tests + private_tests** only (generated_tests omitted). `testtype` is `stdin`; `metadata` is `{}` so the LCB evaluator uses standard input grading. | |
| - This is **not** LiveCodeBench and is **not** an official LCB baseline. The task test set stays `Tommysha/zero-code-l1-lcb-v5-test` (279, 2024-08-01 to 2025-02-01). | |
| - Scoring: `evaluator.py` (LiveCodeBench `testing_util`, 6 s per test, all tests must pass) → pass@1. | |