tracedistill
What it is
A library for distilling complex reasoning chains (CoT) into LoRA adapters using a two-phase 'Train-Nudge' approach. It enforces a strict format contract to ensure models reason step-by-step and output parseable boxed answers.
The problem it solves
Standard SFT often fails to capture the nuances of reasoning traces or can lead to catastrophic forgetting/imbalanced gradients when training on diverse problem types.
What you could use it for
- Distilling reasoning traces from large models into smaller LoRA adapters
- Training models to output specific formats (e.g., \boxed{answer}) for tasks where code execution is not allowed
- Fine-tuning models on complex reasoning tasks like math or logic with type-stratified batching
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.