sokrypton explains why TorchFold self-distillation works

sokrypton explains that running a model thousands of times yields correct high-confidence samples worth adding to training data, citing AF3's jump from 0.3 to 0.6 simply by adding random seeds.

2026-09-18 ~ 2026-09-18 · 2 related posts

• sokrypton explains why self-distillation works for protein structure models — sokrypton · 2026-09-18
• AF3 paper: just increasing seeds lifted scores from 0.3 to 0.6, says sokrypton — sokrypton · 2026-09-18