DatologyAI's gradient-based data filtering cuts jailbreak success by 22%
DatologyAI's new work 'Follow the Gradient' (SafeTulu / SimpleLoRIF) uses scalable gradient attribution instead of static embeddings to filter training data that looks benign but erodes LLM safety, relatively reducing attack success rates by about 22% with little loss in capability.
2026-11-03 ~ 2026-11-03 · 4 related posts
- SimpleLoRIF filters safety-degrading fine-tuning data via gradient alignment — DatologyAI · 2026-11-03
- SafeTulu: scalable gradient attribution filters benign-looking data that erodes LLM safety — pratyushmaini · 2026-11-03
- Datology's SafeTulu: filtering benign-looking data cuts jailbreak success up to 22% — pratyushmaini · 2026-11-03
- DatologyAI's SafeTulu uses gradients to filter unsafe data, cutting attack success 22% — josh_wills · 2026-11-03