RWKV-4 World
Model Description
RWKV-4 trained on 100+ world languages (70% English, 15% multilang, 15% code).
World = Some_Pile + Some_RedPajama + Some_OSCAR + All_Wikipedia + All_ChatGPT_Data_I_can_find
XXXtuned = finetune of World on MC4, OSCAR, wiki, etc.
How to use:
- use https://github.com/josStorer/RWKV-Runner for GUI
- use latest rwkv pip package (0.8.0+)
- use https://github.com/BlinkDL/ChatRWKV/blob/main/v2/benchmark_world.py and https://github.com/BlinkDL/ChatRWKV/blob/main/API_DEMO_WORLD.py to test it
The differences between World & Raven:
- set pipeline = PIPELINE(model, "rwkv_vocab_v20230424") instead of 20B_tokenizer.json (EXACTLY AS WRITTEN HERE. "rwkv_vocab_v20230424" is included in rwkv 0.7.4+)
- use Question/Answer or User/AI or Human/Bot for chat. DO NOT USE Bob/Alice or Q/A
For 0.1/0.4/1.5B models, use fp32 for first layer (will overflow in fp16 at this moment - fixable in future), or bf16 if you have 30xx/40xx GPUs. Example strategy: cuda fp32 *1 -> cuda fp16
NOTE: the new greedy tokenizer (https://github.com/BlinkDL/ChatRWKV/blob/main/tokenizer/rwkv_tokenizer.py) will tokenize '\n\n' as one single token instead of ['\n','\n']
QA prompt (replace \n\n in xxx to \n):
Question: xxx
Answer:
and
Instruction: xxx
Input: xxx
Response:
A good chat prompt (replace \n\n in xxx to \n):
User: hi
Assistant: Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.
User: xxx
Assistant: