reasoning_curation_sampler

What problem does it solve?

It solves unstable training behavior caused by mixing reasoning classes with very different lengths and structures, which leads to gradient variance and poor reasoning transfer.

Core Features & Use Cases

• Conditional Stratification: samples by class first, then applies per-class length filtering to preserve each class’s natural length profile.
• Token Budget Equality: up-weights classes by inverse expected token cost so each class contributes equal token exposure.
• Isomorphic Anchoring: pairs structurally similar but semantically distinct samples to encourage reasoning transfer rather than topical clustering.
• Batch Construction Artifacts: produces a batched stream plus guidance maps for sampling weights, length filters, and curriculum pairing.

Quick Start

Use reasoning_curation_sampler to build a stratified, token-balanced SFT batch stream for a multi-class reasoning dataset with tight per-class length constraints.