Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 28·5 min read
Diff Mining: Logit Differences Reveal Finetuning Objectives
TL;DR
Diff Mining is a novel framework that identifies the learning outcomes of finetuned language models by analyzing logit differences. It effectively reveals salient tokens and finetuning objectives, outperforming existing methods in bias detection and interpretability.
✦ Why It Matters
Implement Diff Mining to audit your finetuned models and uncover hidden biases today.
Key Takeaways
- Use Diff Mining to analyze finetuned model behaviors.
- Implement logit difference analysis for bias detection.
- Explore both Top-K and NMF methods for token aggregation.
- Leverage insights from Diff Mining to improve model interpretability.
How It Works
Diff Mining compares the output logits of a finetuned model with its base model to identify significant changes. It extracts per-context logit differences and aggregates them using methods like Top-K frequency and Non-negative Matrix Factorization (NMF) to create a clear representation of the finetuning objectives.