11 upvotes
LLMs Learn Themselves: Data-Free Self-Distillation Method
What is On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models?
The authors introduce an on-policy self-training technique that lets a language model generate its own training data and fine-tune itself, removing the need for external teachers or curated QA sets. The approach works by having the model answer its own prompts, then using those answers as pseudo-labels for further training.
Who it is for
Researchers and developers who want to boost model performance without gathering new data can try this method to squeeze extra accuracy out of existing LLMs.