deepseek-v4.1POST
Internal safety bench: abliterated DeepSeek V4.1 Flash refuses 2% vs 100% stock
An internal bench compares Morph-abliterated and stock DeepSeek V4.1 Flash: 2% vs 100% refusals with no significant IFEval/MMLU-Pro/GSM8K gaps, and 8% vs 0% recall on whole-repo vuln hunting.
@aaronjmars
@aaronjmarsArchived 4 days agoOriginally posted 2026.10.26
@aaronjmars @aaronjmarsOriginally posted 2026.10.26
from my internal cyber meta bench: refusals on defensive audit + fix tasks (@ArtificialAnlys Cyber Index): Qwen3.8 61-63%, Fable 5.1 16%, GPT-6 Astra 13%, Opus 5.5 9% Morph abliterated vs stock DeepSeek V4.1 Flash: refusals 2% vs 100%, IFEval 87.6 vs 88.7, MMLU-Pro 84.0 vs 83.1, GSM8K 95.1 vs 95.4. none significant, no security tasks our whole-repo vuln hunt: stock V4.1 Flash 8% recall, an uncensored build of it 0% reward hacking: no paired data. DreadIndex caught 21 of 22 stock models cheating
x.com/aaronjmars/status/2114400635592347749Archived 2026.10.26 12:04