brucethemoose's picture
Update README.md
0eccac0
---
license: other
license_name: yi-license
license_link: https://huggingface.co/01-ai/Yi-34B/blob/main/LICENSE
language:
- en
library_name: transformers
pipeline_tag: text-generation
---
Nous-Capybara-34B and Tess-M-Creative-v1.0 merged, then quantized with exllamav2 on 200 rows (400K tokens) on a long Vicuna format chat, a sci fi story and a fantasy story. This should hopefully yield better chat performance than the default wikitext quantization.
Quantized to 4bpw, enough for **~47K context on a 24GB GPU.**
The following merge config was used:
```
models:
- model: /home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied
# no parameters necessary for base model
- model: /home/alpha/Storage/Models/Raw/migtissera_Tess-M-v1.0
parameters:
density: 0.6
weight: 1.0
- model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B
parameters:
density: 0.6
weight: 1.0
merge_method: ties
base_model: //home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied
parameters:
normalize: true
int8_mask: true
dtype: float16
```
First exllama quantization pass:
```
python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K -o /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2 -om /home/alpha/FastModels/capytessmes.json --cal_dataset /home/alpha/Documents/smol.parquet -l 2048 -r 80 -ml 2048 -mr 40 -gr 40 -ss 4096 -nr -b 3.5 -hb 6
```
Second exllama quantization pass:
```
python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K -o /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2 -m /home/alpha/FastModels/capytessmes.json --cal_dataset /home/alpha/Documents/medium.parquet -l 2048 -r 200 -ml 2048 -mr 40 -gr 200 -ss 4096 -b 3.1 -hb 6 -cf /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2-31bpw -nr
```
Both models have Vicuna syntax, so:
# Prompt Format:
```
SYSTEM: ...
USER: ...
ASSISTANT: ...
```
Stop token: `</s>`
***
Credits:
https://github.com/cg123/mergekit
https://huggingface.co/NousResearch/Nous-Capybara-34B/discussions
https://huggingface.co/migtissera/Tess-M-Creative-v1.0
https://huggingface.co/larryvrh/Yi-34B-200K-Llamafied
https://huggingface.co/01-ai/Yi-34B-200K