RealFeedApp Deletion Tuning preserves base-level accuracy
Deletion Tuning achieves results comparable to full fine-tuning by removing fewer than 0.05% of weights, while maintaining prior-task accuracy.
- Field
- training language models
- What they did
- Researchers demonstrated that deleting less than 0.05% of weights in a pretrained large language model allows it to acquire new capabilities, matching full fine-tuning (FFT) across three post-training regimes. They developed the Deletion Tuning method, which learns a binary mask to zero out weights in transformer block gate projections.
- Why it matters
- This method achieves the same results as full fine-tuning while retaining significantly more prior-task accuracy, addressing the central challenge of forgetting in continual learning.
#llm#post-training#deletion tuning#reinforcement learning#continual learning
