# Hugging Face reports 67% accuracy from code-grounded self-taught reasoning traces

> Original title: What worked was skipping the teacher altogether!

- Source: Lewis Tunstall
- Published: 2026-11-02T16:44:16.000Z
- Bad Influence: https://badinfluence.news/items/g39c9csjamb5wo3bwl37996oc
- Original: https://x.com/_lewtun/status/2117296745826631901

## Summary

Lewis Tunstall says his team built training explanations from facts computed in code plus the model's own sense of how unusual a mutation looks, without a teacher model. These grounded reasoning traces raised accuracy to about 67%, while the self-taught version stalled below 60%.

## Post

What worked was skipping the teacher altogether!

After analysing the model errors and prompts more carefully, we realised that each training explanation could be assembled from facts computed in code, plus the model's own sense of how unusual the mutation looks.

With these grounded "reasoning traces", accuracy rose to about 67%. The self-taught version stalled below 60%.

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHWIjvnWWwAAAMv-.png&mode=thumb&exp=1793836800&sig=bf68a25598cce9b0)