Mercan-0.8B-SFT / README.md
Mandotosh's picture
Publish Mercan 0.8B SFT Q4 model
4620cff verified
|
Raw
History Blame
1.34 kB
---
language:
- tr
library_name: mercan
pipeline_tag: text-generation
tags:
- mercan
- nedolm
- gguf
- turkish
- text-generation
license: apache-2.0
---
# Mercan 0.8B SFT
Mercan 0.8B SFT is a Turkish assistant model distributed in the self-contained Mercan v1 deployment format.
## Run
Install the Mercan CLI, then run this repository directly:
```bash
mercan run MercanAI/Mercan-0.8B-SFT
```
The CLI downloads `model.mercan` once into the local Mercan cache and then runs it locally.
You can also download the artifact yourself:
```bash
mercan run model.mercan
```
## Artifact
- `model.mercan`: Q4_K_M deployment model
- Mercan format v1
- physical tensor container: GGUF v3
- exact NDSRF004 tokenizer asset embedded in the model
- context length: 4096
- 24 transformer blocks
- hidden size: 1536
- 12 attention heads / 4 KV heads
- MorphFFN in the first 18 blocks
The `.mercan` file contains model weights, architecture metadata, tokenizer metadata/data and the token-role routing table. It does not contain executable model-bundled code.
## Runtime
The reference runtime and CLI are published in `Ahmet2001/mercanApp-test1`. `libmercan` uses a stable C ABI over the native inference backend.
## Chat template
```text
<|im_start|>{role}\n{content}<|im_end|>\n
```
EOS is appended once at the end of a training conversation.