sanchitahuja205/xelm-gemma-4b-indic-window-10-16-freeze
Partial fine-tune (fewer layers) vs google/gemma-3-4b-pt. sanchitahuja205/xelm-gemma-4b-indic-window-10-16-freeze matches part of google/gemma-3-4b-pt (444 of its 883 tensors; the rest of google/gemma-3-4b-pt has no counterpart). On the shared part: Partial change: 85% of the weights changed (attention, embeddings, mlp, norms), in 28 of 34 layers.
Tensors changed: Embeddings 1/1, Norms 67/205, MLP 84/102, Attention 112/136