final_sft_corrected / README.md
Mandotosh's picture
Add renamed final Mercan artifact: Mercan-0.8B-Final-Q4_K_M.mercan
d57289f verified
|
Raw History Blame Contribute Delete
1.18 kB
---
language:
- tr
library_name: gguf
tags:
- mercan
- nedolm
- text-generation
- q4_k_m
- turkish
---
# Mercan 0.8B — Final SFT + DPO + GRPO/RLVR (Q4_K_M)
Final Mercan v1 deployment artifact for the NedoLM/Mercan 0.8B Turkish assistant.
## Files
- `Mercan-0.8B-Final-Q4_K_M.mercan`: GGUF-v3 physical container, Mercan v1 metadata, Q4_K_M quantization
- `Mercan-0.8B-Final-Q4_K_M.mercan.sha256`: checksum
- `Mercan-0.8B-Final-Q4_K_M.mercan.json`: provenance/conversion manifest
Source chain: **continuation SFT -> DPO -> GRPO/RLVR**. Final source checkpoint: `step_00002064.pt`.
Model contract: NDSRF004 tokenizer (SHA-256 `72412d981dac65a29d1767bc98821fc2bcffc2de53c534e7c719598515bfb600`), 32,002 model vocab rows, control IDs 32000/32001, canonical Turkish roles `sistem`, `kullanici`, `asistan`, hidden size 1536, 24 layers, 12 attention heads, 4 KV heads, FFN 5632, context/sliding window 32768, RoPE base 1,000,000, 18 MorphFFN layers.
The structural control IDs are exported with Mercan runtime aliases `<|im_start|>` and `<|im_end|>`; the trained model consumes the numeric control IDs. Role headers remain Turkish.
```bash
mercan run MercanAI/final_sft_corrected
```