|
--- |
|
inference: false |
|
language: |
|
- zh |
|
- en |
|
license: unknown |
|
model_name: Chihiro-7B-v0.1 |
|
pipeline_tag: text-generation |
|
prompt_template: '<s> SYS_PROMPT [INST] QUERY1 [/INST] RESPONSE1 [INST] QUERY2 [/INST]' |
|
tags: |
|
- nlp |
|
- chinese |
|
- mistral |
|
- traditional_chinese |
|
- merge |
|
- mergekit |
|
- MediaTek-Research/Breeze-7B-Instruct-v0_1 |
|
- mlabonne/Zebrafish-7B |
|
--- |
|
|
|
<br/> |
|
|
|
# 千尋 7B v0.1 |
|
|
|
Zebrafish 7B 加上 Breeze 7B 的 slerp merge 試驗性通用繁中基座模型 📚 |
|
|
|
GGUF Quants 👉 [Chihiro-7B-v0.1-GGUF](https://huggingface.co/yuuko-eth/Chihiro-7B-v0.1-GGUF) |
|
|
|
請用 Mistral 7B Instruct 或是 Breeze 7B Instruct 所推薦的 Prompt 格式進行操作；以下為模型配置。 |
|
|
|
 |
|
|
|
### Chihiro 7B v0.1 |
|
|
|
This is an experimental Mistral-architecture SLERP merge with two brilliant base models. Zebrafish and Breeze were used together in this work. |
|
|
|
Model configuration is as follows: |
|
|
|
* [Breeze-7B-Instruct](https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v0_1) as base. |
|
* [Zebrafish-7B](https://huggingface.co/mlabonne/Zebrafish-7B) as model 1. |
|
|
|
To use the model, please use either prompt templates suggested by the base models, or just slap the Mistral one on. |
|
|
|
<br/><br/> |
|
|
|
### Benchmarks |
|
|
|
Evaluation suite: OpenLLM |
|
| Model | ARC |HellaSwag| MMLU |TruthfulQA|Winogrande|GSM8K| |
|
|-------------------------------------------------------------------|----:|--------:|--------------------------|---------:|---------:|----:| |
|
|[Chihiro-7B-v0.1](https://huggingface.co/yuuko-eth/Chihiro-7B-v0.1)|68.52| 85.95| (not yet evaluated) | 63.81| 81.77|64.22| |
|
|
|
|
|
Evaluation suite: Nous |
|
| Model |AGIEval|GPT4All|TruthfulQA|Bigbench|Average| |
|
|-------------------------------------------------------------------|------:|------:|---------:|-------:|------:| |
|
|[Chihiro-7B-v0.1](https://huggingface.co/yuuko-eth/Chihiro-7B-v0.1)| 45.16| 75.26| 63.82| 47.38| 57.91| |
|
|
|
|
|
Average: 47.38% |
|
|
|
Average score: 57.91% |
|
|
|
Evaluated Apr. 27, 2024, NVIDIA RTX 4090 |
|
|
|
<br/><br/> |
|
|
