Summary
- Transluce researchers trained oversight models of varying sizes to predict how finetuning changes a subject model's behavior based on the finetuning dataset.
- Results show that performance follows a smooth power law as training data and model size increase, except for a privileged access effect where an oversight model matching the subject model outperforms larger ones.
- Experiments used Qwen3.5-2B as the base subject model and trained oversight models across various sizes using datasets derived from FineFineWeb documents.
- The study observed an apparent performance floor where validation loss asymptotes at non-zero levels even with increased compute, likely reflecting response variance.
Why it matters
Predicting the side effects of model finetuning is crucial for safety oversight, helping developers anticipate unexpected or misaligned behaviors before deployment.