rwkv-4-pile-3b / README.md
BlinkDL's picture
Update README.md
ecb46be
|
raw
history blame
948 Bytes
metadata
language:
  - en
tags:
  - pytorch
  - text-generation
  - causal-lm
  - rwkv
license: apache-2.0
datasets:
  - The Pile

RWKV-4 3B

Model Description

RWKV-4 3B is a L32-D2560 causal language model trained on the Pile. See https://github.com/BlinkDL/RWKV-LM for details.

** Note: It's a BF16 model, and it may overflow if you are using FP16 (probably fixable by rescaling the weights). **

At this moment you have to use my Github code (https://github.com/BlinkDL/RWKV-LM) to run it.

ctx_len = 1024 n_layer = 32 n_embd = 2560

Preview checkpoint: RWKV-4-Pile-3B-20220921-3047.pth : Trained on the Pile for 125B tokens.

  • Pile loss 2.0026
  • LAMBADA ppl 5.72, acc 61.36%
  • PIQA acc 73.39%
  • SC2016 acc 68.84%
  • Hellaswag acc_norm 56.57%

Preview checkpoint: RWKV-4-Pile-3B-20220915-1207.pth : Trained on the Pile for 50B tokens.

  • Pile loss 2.0902
  • LAMBADA ppl 7.01, acc 57.11%
  • PIQA acc 72.52%
  • SC2016 acc 68.36%
  • Hellaswag acc_norm 52.17%