Datasetpublic

baesad/s1K-DeepSeek-R1-Qwen-32B-evaluated

s1K DeepSeek R1 Distill Qwen 32B — evaluated This dataset preserves the 1,000 rows and original columns from VoCuc/s1K-1.1-DeepSeek-R1-Distill-Qwen-32B and adds correctness annotations for generated_response against solution. Added columns is_correct: whether the generated answer is judged correct. evaluation_method: math_verify_numeric_visible_response for a reference that is exactly one numeric literal, otherwise manual_visible_response_review.… See the full description on the dataset page: https://huggingface.co/datasets/baesad/s1K-DeepSeek-R1-Qwen-32B-evaluated.

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes82downloads

Dataset Card

s1K DeepSeek R1 Distill Qwen 32B — evaluated

This dataset preserves the 1,000 rows and original columns from `VoCuc/s1K-1.1-DeepSeek-R1-Distill-Qwen-32B` and adds correctness annotations for generated_response against solution.

Added columns

—is_correct: whether the generated answer is judged correct.
—evaluation_method: math_verify_numeric_visible_response for a reference that is exactly one numeric literal, otherwise manual_visible_response_review.
—evaluation_reason: extracted math-verify values or a concise manual-review reason.

Evaluation

—Total: 1,000
—Correct: 559 (55.9%)
—Incorrect: 441 (44.1%)
—Numeric-literal solution verified with math-verify: 336
—All other solution values manually reviewed: 664

Only text after the first </think> marker (or the whole response if there is no marker) is evaluated. Math verification is intentionally restricted to a reference solution matching a single numeric literal, e.g. 123 or 12.5. Every other sample was read against its question and reference solution. The annotations are auditable through evaluation_method and evaluation_reason.