---
tags: training
---
# self-play-finetune

Fine-tune the prover on its own successful proofs.

## What happened

- 2026-10-02: 82 problems proved (+18 on the baseline)

## Next

Small gains for the cost; needs more data.