Hugging Face's Lewis Tunstall reports 67% accuracy from grounded self-taught reasoning traces

Lewis Tunstall says his team dropped the teacher model and built training explanations from facts computed in code, plus the model's own estimate of how unusual a mutation looks.

He reports these grounded reasoning traces raised accuracy to about 67%, while the self-taught version stalled below 60%. The source is a single post on X, and the post gives no benchmark name or full method.