c-bone/CrystaLLM-pi_Mattergen-XRD
Full fine-tune (fewer layers) vs c-bone/CrystaLLM-pi_base. c-bone/CrystaLLM-pi_base matches part of c-bone/CrystaLLM-pi_Mattergen-XRD (100 of its 164 tensors; the rest of c-bone/CrystaLLM-pi_Mattergen-XRD has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 8 of 8 layers.
Tensors changed: Attention 32/32, Norms 34/34, MLP 32/32, Embeddings 2/2