finding
active
finding:dpo-on-dolci-yields-lower-aggregate-safety-score-than-sft-checkpoint-e-g-llama-3-1-8b-0-849-to-0-758DPO on Dolci yields lower aggregate safety score than SFT checkpoint (e.g. Llama 3.1 8B 0.849 to 0.758)
Safeguard degradation.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Egregious behavior discovered.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Benchmark identification rate rises from 0.08 (SFT) to 0.25 (DPO); interventions barely reduce itfinding0.768Table 11.
- App. B.
- K-FAC edit reduces OLMo-7B Dolma strict accuracy from 99.9% to 3.4% (BSN 6.0%, SVD 3.0%)finding0.742Table 1.
- Demonstrates the paradox magnitude relative to semantically dangerous personas
- Qualitative failure mode difference between architectures under activation steering
- Key observation that SP rankings are preserved cross-architecturally while AS is not
- Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
- Strongest probe validation result; highest Cohen's d among the four concepts