NedoLM-0.8B-SFT-GGUF

Public GGUF snapshot of NedoLM 0.8B after Turkish full-parameter SFT.

• SFT checkpoint: step_00005300.pt
• Base pretraining checkpoint: /arf/scratch/egitimg16/nedolm_0p8b/runs/nedolm-0p8b-v1/checkpoints/step_00064000.pt
• Parameters: 823,076,352
• Context length: 4096
• Attention: 12 heads / 4 KV heads, head dim 128
• Sliding window: 2048
• FFN width: 5632
• MorphFFN layers: 18 / 24
• GGUF storage: F16
• GGUF SHA256: 9c830384e5434035e77c55ae40c5940b386f42156b63c7d4132af5b2ec986aa1
• NedoTokenizer vocab SHA256: 72412d981dac65a29d1767bc98821fc2bcffc2de53c534e7c719598515bfb600

Important runtime note

This file uses the custom GGUF architecture name nedolm. NedoLM contains token-prior MorphFFN routing (OTHER/ROOT/SUFFIX channels). Stock llama.cpp does not understand this architecture unless a NedoLM/MorphFFN runtime implementation is added. The GGUF is a valid v3 container and preserves the model tensors and architecture metadata; surface-vocab.bin is included for the exact tokenizer used by training.

SFT data

The SFT loader trains only on assistant targets and excludes conversations containing tool-role messages. During the corpus audit, 3,753 tool-role documents were detected and filtered from the readable training corpus.