YYYYYYibo/VSR-MATH500-Qwen2.5-7B-Rollouts

VSR MATH-500 Qwen2.5-7B Rollouts

This dataset contains 8,000 source trajectories sampled from Qwen/Qwen2.5-7B-Instruct on the 500 problems in HuggingFaceH4/MATH-500 (16 trajectories per problem).

The user prompt appends:

Solve the problem with a detailed justification. Put only the final answer inside \boxed{}.

Sampling

• —Samples per problem: 16
• —Temperature: 0.7
• —Top-p: 0.95
• —Maximum new tokens: 8,192
• —Base seed: 20260730
• —Chat template: the official Qwen2.5-7B-Instruct tokenizer template

Contents

math500_qwen25_7b_sources_8k.jsonl is append-only source data with one rollout per line. Each record includes:

• —the MATH-500 problem ID, problem text, and reference answer;
• —rendered prompt text and exact prompt token IDs;
• —generated response text and exact response token IDs;
• —token and character counts;
• —sampling configuration and termination reason;
• —the final extracted \boxed{} answer;
• —a conservative normalized-string-match correctness label.

All 500 problems have exactly 16 records. There are 8,000 unique (problem_id, sample_index) pairs.

Important label limitation

is_correct_string_match is a preliminary exact-match label after conservative textual normalization. It is not a symbolic mathematics grader and can mark mathematically equivalent expressions as different. Consumers should re-grade the raw responses for evaluations requiring reliable mathematical equivalence.

Of the 8,000 trajectories, 17 reached the 8,192-token generation limit and 31 did not contain an extractable \boxed{} expression.

Intended use

The trajectories were collected for controlled analysis of continuation success from intermediate response prefixes, including comparisons between prefixes of successful and failed source trajectories.