Dataflow Pattern Math Reasoning Sft

Math Reasoning SFT Pattern

Use this pattern for short, verifiable math problems whose training output requires a standalone instruction, regenerated reasoning, and a final answer compatible with strict benchmark evaluation. Do not use it for proof-only tasks, open-ended tutoring, or records without a trustworthy answer signal.

The recommended case is gsm8k-reasoning-v1. It accepts question/answer and problem/final_answer schema branches, normalizes them to canonical fields, regenerates reasoning without exposing the gold answer to the generator, validates the derived final answer against gold, and applies an LLM quality evaluation after deterministic correctness checks.

Preserve these invariants when adapting the case:

• sample_id and source dataset provenance survive every stage.
• The reasoning generator sees the problem but not the gold answer.
• Generated reasoning is accepted only when its parsed conclusion matches gold.
• Quality thresholds are justified by the trial score distribution.
• Output always provides nonempty instruction, reasoning, and output.
• Benchmark answer formatting is handled explicitly rather than left to chance.

For a new schema or benchmark, retain the staged validation pattern but adapt normalization, correctness parsing, difficulty coverage, and final formatting to the actual task.