PHLoRA extracts LoRA adapters from fine-tuned models

Published on September 17, 2025, the post describes PHLoRA, a method that uses base and fine-tuned checkpoints to extract a low-rank LoRA adapter through singular value decomposition. The reported process requires neither training data nor gradients.

The post says accuracy at ranks 32 or 64 is within about 1% of full-rank models, and notes potential reductions in adapter loading and serving costs when converting existing full-rank fine-tunes. To check the scope, method, and findings, consult the original post and verify its reported conditions and metrics.