rwkv-4-pile-3b / README.md
BlinkDL's picture
Update README.md
bd0d392
|
raw
history blame
964 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.

New checkpoint: RWKV-4-Pile-3B-20221110-ctx4096.pth : Fine-tuned to ctx_len = 4096

  • LAMBADA ppl 5.25, acc 63.96%
  • PIQA acc 74.16%
  • SC2016 acc 70.71%
  • Hellaswag acc_norm 59.89%
  • ctx_len = 4096 n_layer = 32 n_embd = 2560

Final checkpoint: RWKV-4-Pile-3B-20221008-8023.pth : Trained on the Pile for 331B tokens.

  • Pile loss 1.9469
  • LAMBADA ppl 5.24, acc 63.94%
  • PIQA acc 73.72%
  • SC2016 acc 70.28%
  • Hellaswag acc_norm 59.63%
  • ctx_len = 1024 n_layer = 32 n_embd = 2560