Narumashi-11B-v0.9 / README.md
Alsebay's picture
Update README.md
adc99d7 verified
---
language:
- en
license: cc-by-nc-4.0
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- trl
- sft
- Roleplay
- roleplay
base_model: Sao10K/Fimbulvetr-11B-v2
---
# About this model
This model can handle (limited) TSF content. If you Character Card have complex plot, maybe you should try other model (maybe bigger parameter?).
- Early stop version of https://huggingface.co/Alsebay/Narumashi-11B , expect to no overfitting when finetune, maybe good ?
Do you know TSF, TS, TG? A lot of model don't really know about that, so I do some experiment to finetune TSF dataset.
- **Finetuned with Chinese Novels dataset, to increase the accuracy in TSF theme, which is not quite popular.
(R18 dataset). You should include chinese/japanese word about tag you want(search it in pixiv) in your character card to trigger it.
This finetune idea is suitable for Chinese Roleplay than English (Becaue I could only find good Chinese datasets about it πŸ™ƒ, it is nice that if you can open a discussion about English TSF datasets). But it still affect the models writing styles, so maybe less GPT-like response in both Chinese and English?.**
- **Finetuned from model :** Sao10K/Fimbulvetr-11B-v2 . Thank Sao10K a lot :)
## 8k Context Length BTW, the original Fimbulvetr and Solar have only 4k context length, so I extended it πŸ˜†.
## GGUF version? [here is it](https://huggingface.co/Alsebay/Narumashi-11B-v0.9-GGUF).
## Dataset
All chinese novels dataset
```
Dataset(all are novels):
60% skinsuit
25% possession
5% transform(shapeshift)
10% other
```
# Thank Unsloth for good finetuning tool. This mistral model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)