Sakalti/SJTPass-4
Partial fine-tune (fewer layers) vs Sakalti/SJT-0.5B. Sakalti/SJT-0.5B matches part of Sakalti/SJTPass-4 (291 of its 723 tensors; the rest of Sakalti/SJTPass-4 has no counterpart). On the shared part: Partial change: 28% of the weights changed (attention, mlp, norms), in 12 of 24 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 24/49, MLP 36/72, Attention 84/168
1 derivatives checked
- mradermacher/SJTPass-4-GGUF: Same weights, converted