|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: |
|
|
- RWKV-Red-Team/ARWKV-7B-Preview-0.1 |
|
|
--- |
|
|
|
|
|
This model only works with RWKV-Infer and RWKV-LM-RLHF. |
|
|
This is a completely early model. No warranty. Use at your own risk. |
|
|
|
|
|
slightly fine-tuned CJE 900k including Deepseek R1 magpie. |
|
|
|
|
|
v7 FLA ctx5120, 1k slided overlap training. SFT(with loss smooting=0.001) |
|
|
DoRA(total 0.5B parameters), Emb frozen. |
|
|
|
|
|
if you want use reasoning mode, set system prompts below |
|
|
|
|
|
``` |
|
|
Answer user questions in Reasoning mode. |
|
|
``` |
|
|
|
|
|
for using |
|
|
1. install RWKV-Infer(see how to install) |
|
|
2. loadmodel(choose fp16 or fp6 or fp5 (dont choose FP8)) |
|
|
3. need 19GB VRAM in FP16, 12GB VRAM in FP6 |
|
|
``` |
|
|
curl http://127.0.0.1:9000/loadmodel -X POST -H "Content-Type: application/json" -d '{"model_filename":"models/ARWKV-7B-CJE-30%.pth","model_viewname":"ARWKV-7B-CJE-30%","model_strategy":"fp16"}' |
|
|
|
|
|
``` |
|
|
3. you can use this model via OpenAI CompatibleAPI http://127.0.0.1:9000/v1 and set modelname "ARWKV-7B-CJE-30%" |
|
|
4. its good for sillytavern |
