rwkv-4-pile-1b5 / README.md
BlinkDL's picture
Update README.md
933876b
|
raw
history blame
1.25 kB
metadata
language:
  - en
tags:
  - pytorch
  - text-generation
  - causal-lm
  - rwkv
license: apache-2.0
datasets:
  - The Pile

RWKV-4 1.5B

Model Description

RWKV-4 1.5B is a L24-D2048 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 = 24 n_embd = 2048

New checkpoint: RWKV-4-Pile-1B5-20220929-ctx4096.pth : Fine-tuned to ctx_len = 4096

Final checkpoint: RWKV-4-Pile-1B5-20220903-8040.pth : Trained on the Pile for 332B tokens.

  • Pile loss 2.0415
  • LAMBADA ppl 7.04, acc 56.43%
  • PIQA acc 72.36%
  • SC2016 acc 68.73%
  • Hellaswag acc_norm 52.48%

Preview checkpoint: RWKV-4-Pile-1B5-20220822-5809.pth : Trained on the Pile for 240B tokens.

  • Pile loss 2.0518
  • LAMBADA ppl 7.14, acc 56.36%
  • PIQA acc 71.71%
  • SC2016 acc 68.15%
  • Hellaswag acc_norm 52.04%

Preview checkpoint: RWKV-4-Pile-1B5-20220814-4526.pth : Trained on the Pile for 187B tokens.

  • Pile loss 2.0635
  • LAMBADA ppl 7.34, acc 55.64%
  • PIQA acc 71.44%
  • SC2016 acc 68.25%
  • Hellaswag acc_norm 51.60%