DeepSeek's Hidden Bias: How We Cut It by 76% Without Performance Loss 1 year ago
reach out at @nicilevv on X for questions
HN user
nicolevin
52 karma
Posts1
Comments3
Bias-Unlearned DeepSeek-R1-Distill-Llama-8B here: https://huggingface.co/hirundo-io/DeepSeek-R1-Distill-Llama-...
DeepSeek-R1 (8B) exhibited 2x more bias than base Llama. We applied targeted unlearning, reduced bias by up to 76% across race/gender/nationality, while maintaining model performance (TruthfulQA: 9.8→9.9, LogiQA: 42.6%→42.5%). Done in ~1hr on consumer hardware. Debiased model on HuggingFace.