Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 3·5 min read
The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models
TL;DR
Knowledge distillation in small language models can reduce bias in clear tasks but exacerbate it in ambiguous ones. For instance, it improved context-following in SmolLM2-1.7B-Instruct but led to stereotype responses in 15% of ambiguous cases.
✦ Why It Matters
Evaluate your training data to mitigate bias in language models, especially in ambiguous contexts.
Key Takeaways
- Implement knowledge distillation cautiously in ambiguous tasks.
- Monitor context-following errors in language models.
- Audit training data to identify bias sources.
- Consider distinct mechanisms behind bias effects.
How It Works
Knowledge distillation allows a smaller model to learn from a larger one, improving its performance on specific tasks. In this study, the Gemma-2-9B teacher model provided responses that the smaller models used to refine their context-following abilities.
However, the process can inadvertently introduce biases, particularly in ambiguous scenarios where the model's decision-making is less clear.