Effectiveness and durability of pretraining data filtering in protein language models

The researchers evaluate pretraining data filtering as a safety defense on the ESM-2 family of protein language models (8M to 3B parameters), finding that filtering viral proteins from pretraining data significantly reduces the model's masked language modeling, mutation effect prediction, and long-range contact prediction performance on viral proteins, while largely preserving performance on non-viral proteins. These effects are concentrated on proteins of viruses that can infect humans and increase with model scale, with stricter filtering being more effective on larger models. However, the lost performance can be recovered by fine-tuning on the filtered viral family sequences, with recovery cost rising with scale, and even the largest model requires only about $80 of H100 machine time. The authors therefore conclude that pretraining filtering can moderately and selectively weaken viral protein-related capabilities, but it is not durable against fine-tuning that requires only a moderate amount of compute and relevant data.