|
--- |
|
license: apache-2.0 |
|
--- |
|
|
|
This is full finetuned model from RWKV 4 world 7B CHNTuned model using data from Readflow tech (readflow.com.cn) , |
|
|
|
finetuned for 32k context length, used to summary news article |
|
|
|
using inf-ctx training https://github.com/SynthiaDL/TrainChatGalRWKV with fixed VRAM |
|
|
|
you can test summary prompt using RWKV runner(https://github.com/josStorer/RWKV-Runner) in chat mode , and check conversation files in examples folders. |
|
|
|
https://discord.gg/pWH5MkvtNR |
|
|
|
![QQ图片20230721210758.jpg](https://cdn-uploads.huggingface.co/production/uploads/6176b32847ee6431f632981e/NTa346CoQzH-k0civqG1Q.jpeg) |
|
|
|
------------------------------------------------------------------------------------------------- |
|
|
|
|
|
这是和会读Readflow合作的模型( readflow.com.cn),用于超长微信文章的摘要,训练了32k的context长度,能一次性输入整篇文章进行摘要,如下图以及在example文件夹中的例子, |
|
|
|
23k tokens一次性输入摘要,不增显存,如果有state状态,可以秒出,可以反复重试多种可能性, |
|
|
|
“Assistant:"是模型输出内容 |
|
|
|
|
|
可以使用runner进行测试 https://github.com/josStorer/RWKV-Runner |
|
|
|
RWKV world的token对中文以及各类语言效率很高,token和字比例,基本是1:1,甚至1:n,欢迎测试,并加入微调QQ群讨论439087067 |
|
|
|
|
|
|
|
![微信截图_20230720145202.png](https://cdn-uploads.huggingface.co/production/uploads/6176b32847ee6431f632981e/3prZb97XOKeHhVQyYevA_.png) |
|
|
|
|
|
|