TraceForge writes student-style reasoning traces for distillation
Original title: TraceForge: Distilling with Your Own Thinking
With the same 5,000 questions and teacher answers, HumanEval+ rises 3.7 points and MBPP+ 7.1.
TraceForge generates a reasoning trace from context and a teacher's answer in the student model's family style, after some frontier teams stopped returning detailed traces. Training is supervised then adversarial: a generator rebuilds family traces and a critic scores how real a trace looks. An independent judge rates generated traces 0.68 authentic versus 0.66 for real traces. On coding, the generated traces beat Fable 5's own thinking; general-task corpora are about the same.
Unconfirmed: The authors suspect family-style reasoning is easier to learn from, but that is not established.
- Critic-selected traces from four candidates score 0.82 authenticity.
- Those figures are authenticity scores, not correctness rates.
- The MBPP+ lead holds at every checkpoint.
- The method avoids replaying encrypted blocks that break provider terms.
From the original article
1 min read at huggingface.co
Using an LLM to generate training data for another is among the most general ways to train one, and it remains common practice. Some frontier teams have closed that route for their own models: what comes back is an answer, or a short summary, without the detailed reasoning trace behind it.
By Yan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.