metadata

language:
  - en
  - zh
  - de
  - fr
  - es
  - pt
  - ru
  - it
  - ja
  - ko
  - vi
  - ar
tags:
  - pytorch
  - text-generation
  - causal-lm
  - rwkv
license: apache-2.0
datasets:
  - EleutherAI/pile
  - togethercomputer/RedPajama-Data-1T

RWKV-5 World (Training in Progress)

Note https://huggingface.co/BlinkDL/rwkv-4-world are the best models as of now, because I am still training RWKV-5.

Use rwkv pip package 0.8.7+ for RWKV-5 inference.

Model Description

RWKV-5 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

RWKV-5 training: set --my_testing "r" in latest RWKV-LM

RWKV-5 inference implementation reference: https://github.com/BlinkDL/ChatRWKV/blob/main/RWKV_v5_demo.py

0.1B world:

RWKV-5 37% trained = LAMBADA ppl 18.1 acc 42.93%

RWKV-4 100% trained = LAMBADA ppl 25.5 acc 36.29%