On Oct 26, UK AISI: pretraining data filters on open-weight bio models work but fine-tuning can restore capability.
On Oct 26, UK AISI: pretraining data filters on open-weight bio models work but fine-tuning can restore capability.
UK AISI published this on Oct 26, 2026 (source: aisi.gov.uk).
The UK AI Security Institute published a study on pretraining data filtering as a safeguard for open-weight protein models, finding filters can be effective and selective but not necessarily durable against fine-tuning. In one result, an ESM-2 model with viral proteins removed fully recovered capability after fine-tuning; a 3B model needed about $80 of compute.
Items older than 72 hours never go in Top stories and show their original date.