Sakalti/SJT-900M
Partial fine-tune (fewer layers) vs Sakalti/SJT-0.5B. Sakalti/SJT-0.5B matches part of Sakalti/SJT-900M (291 of its 507 tensors; the rest of Sakalti/SJT-900M has no counterpart). On the shared part: Partial change: 43% of the weights changed (attention, mlp, norms), in 18 of 24 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 36/49, MLP 54/72, Attention 126/168
2 derivatives checked
- mradermacher/SJT-900M-GGUF: Same weights, converted
- Sakalti/SJT-900M-Alpha: Same weights, repacked