Sakalti/SJTpass-3
Partial fine-tune (fewer layers) vs Sakalti/SJT-0.5B. Sakalti/SJT-0.5B matches part of Sakalti/SJTpass-3 (291 of its 483 tensors; the rest of Sakalti/SJTpass-3 has no counterpart). On the shared part: Partial change: 38% of the weights changed (attention, mlp, norms), in 16 of 24 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 32/49, MLP 48/72, Attention 112/168
1 derivatives checked
- mradermacher/SJTpass-3-GGUF: Same weights, converted