ar0cket1/hintedselfteacher-nemotron-math-v2-AoPS

hintedselfteacher-nemotron-math-v2-AoPS

This dataset contains a training-ready hinted self-teacher split derived from the AoPS split of nvidia/Nemotron-Math-v2.

The source problems were filtered to the AoPS split with the medium/notool solve rate between 2 and 6. Hints were generated with GPT-5.5 medium using an h17_nt hint-generation prompt. This hint type was close to the best hint type found after doing hint mutations, based on qualitative analysis of token-level hinted self-teacher behavior.

Contents

• —data/train.parquet: canonical training file.
• —data/train.jsonl: JSONL mirror for inspection or non-Parquet loaders.
• —manifest.json: schema and intended training-use metadata.
• —sample.json: a few example rows.

Rows: 29,315.

Row Schema

Important columns:

• —problem_id: stable problem id.
• —problem: raw student problem text, without private hint.
• —verifier_answer: final verifiable answer/reference string.
• —normalized_verifier_answer: normalized answer string for lightweight checks.
• —hint_id: always h17_nt.
• —hint_text: private hinted self-teacher context generated for the problem.
• —student_messages: chat-format rollout prompt for the student.
• —teacher_messages: chat-format hinted teacher prompt for teacher-forced scoring.
• —student_user_content: flattened student prompt.
• —teacher_user_content: flattened hinted teacher prompt.
• —teacher_problem_text: exact privileged-wrapper text used before chat templating / answer-instruction append.

The export intentionally excludes worked solutions from the training rows. It contains the final verifiable answer and generated hint, but not the official solution chain.

Intended Use

For hinted self-teacher / OPSD-style experiments:

1. 1.Generate rollouts from student_messages.
2. 2.Score the same sampled tokens under teacher_messages.
3. 3.Use sampled-token pressure logp_teacher(token | teacher context + prefix) - logp_student(token | student context + prefix).
4. 4.Use verifier_answer for final-answer reward or filtering.

Provenance

• —Source: nvidia/Nemotron-Math-v2.
• —Subset: AoPS split.
• —Filter: medium/notool solve rate between 2 and 6.
• —Hint generator: GPT-5.5 medium.
• —Hint family: h17_nt, selected after hint-mutation experiments as a strong qualitative hinted self-teacher prompt.