Aletheia: Gradient-Guided Layer Selection for Efficient LoRA Fine-Tuning Across Architectures

Build an 'Aletheia' trainer for HuggingFace PEFT. It should automatically prune LoRA layers that aren't learning, reducing VRAM usage.

Suggested repo: aletheia-lora

"Stop wasting VRAM on ineffective LoRA layers."

Estimated effort: 50h