CRMA Continual learning

Been experimenting with CRMA — a stability layer for LoRA fine-tuning that handles sequential training without catastrophic forgetting. Tested on TinyLlama and Mistral-7B. Full results dropping soon.

Thanks for this post.