rwkv-4-raven / README.md
BlinkDL's picture
Update README.md
287c586
|
raw
history blame
No virus
1.25 kB
---
language:
- en
tags:
- pytorch
- text-generation
- causal-lm
- rwkv
license: apache-2.0
datasets:
- the_pile
---
# RWKV-4 "Raven"-series Models
## Model Description
These are RWKV-4-Pile models 3B/7B/14B finetuned on Alpaca, CodeAlpaca, Guanaco, GPT4All, ShareGPT and more. Ctxlen 8192 in v9.
Gradio Demo: https://huggingface.co/spaces/BlinkDL/Raven-RWKV-7B
Use https://github.com/BlinkDL/ChatRWKV to run them.
See https://github.com/BlinkDL/RWKV-LM for details on the RWKV Language Model (100% RNN).
Best Prompt Format for Raven models, Bob is user, Alice is bot (NOTE: no space after final "Alice:"). You can use \n within xxxxxxxxxxx, but avoid \n\n.
```
Bob: xxxxxxxxxxxxxxxxxx\n\nAlice:
Bob: xxxxxxxxxxxxxxxxxx\n\nAlice: xxxxxxxxxxxxx\n\nBob: xxxxxxxxxxxxxxxx\n\nAlice:
```
New models will be named like Eng99%-Other1%, Eng86%-Chn10%-JpnEspKor2%-Other2%, etc.
Language ratios determined by amount of ChatGPT data. Please share more ChatGPT data to increase the ratio of your language.
Old models:
* RWKV-4-Raven-Eng : 99% English + 1% Multilang
* RWKV-4-Raven-EngAndMore : 96% English + 2% Chn Jpn + 2% Multilang (More Jpn than v6 "EngChnJpn")
* RWKV-4-Raven-ChnEng : 49% English + 50% Chinese + 1% Multilang
License: Apache 2.0