NedoLM-0.8B-SFT-GGUF

GGUF family for NedoLM 0.8B Turkish SFT, exported from SFT checkpoint step_00005300.pt.

Files

Variant	Size	Suggested use
Q8_0	0.82 GiB	Highest-fidelity quantized release
Q4_0	0.43 GiB	Compact local/storage-oriented release
F16	1.53 GiB	Reference full-precision GGUF

Exact tokenizer asset surface-vocab.bin is included. It is NedoTokenizer NDSRF004, vocabulary size 32,000, SHA256 72412d981dac65a29d1767bc98821fc2bcffc2de53c534e7c719598515bfb600.

Architecture

• 823M parameters
• 24 decoder blocks
• d_model 1536
• 12 attention heads / 4 KV heads
• context 4096
• sliding window 2048
• 18/24 TokenPrior MorphFFN layers
• tied token embeddings

Quantization

The 2-D weight tensors are block-quantized with GGML-compatible Q8_0 or Q4_0 block layouts. 1-D norm and routing tensors remain F16. structural_validation.json records sizes, hashes, tensor counts and tensor-type counts.

Runtime note

NedoLM uses the custom GGUF architecture name nedolm and token-prior MorphFFN routing. Stock llama.cpp does not currently implement this architecture. These GGUF files preserve the model and standardized GGML tensor quantization formats, but inference requires a NedoLM/MorphFFN runtime implementation.

Training note

SFT used assistant-only loss. A strict audit found 3,753 residual tool-role documents in the source corpus; the loader excluded those complete conversations from training.