license: apache-2.0
datasets:
- Norquinal/claude_multiround_chat_30k
- OpenLeecher/Teatime
We proudly announce this is the world first 128k context model based on RWKV architecture today, 2023-08-10.
With RWKV world tokenizer,multi-langs have 1:1 tokenization ratio ,one word to one token. (https://github.com/BlinkDL/ChatRWKV/blob/2a13ddecd81f8fd615b6da3a8f1091a594689e30/tokenizer/rwkv_tokenizer.py#L163)
This model trained with instructions datasets and chinese web novel and tradition wuxia, more trainning details would be updated.
Tested to summary 85k tokens to 5 keypoints ,can find conversation files in example folders ,more cases are coming.
Full finetuned using this repo to train 128k context model , 4*A800 40hours with 1.3B tokens. https://github.com/SynthiaDL/TrainChatGalRWKV/blob/main/train_world.sh
Using RWKV Runner https://github.com/josStorer/RWKV-Runner to test this , use temp 0.1-0.2 topp 0.7 for more precise answer ,temp between 1-2.x is more creatively.