rwkv-4-pile-14b / README.md
BlinkDL's picture
Update README.md
7a2e9f4
|
raw
history blame
789 Bytes
metadata
language:
  - en
tags:
  - pytorch
  - text-generation
  - causal-lm
  - rwkv
license: apache-2.0
datasets:
  - the_pile

RWKV-4 14B

Model Description

RWKV-4 14B is a L40-D5120 causal language model trained on the Pile. See https://github.com/BlinkDL/RWKV-LM for details.

Use https://github.com/BlinkDL/ChatRWKV to run it.

ctx_len = 1024 n_layer = 40 n_embd = 5120

RWKV-4-Pile-14B-2023xxxx-ctx4096-testxxx.pth : Fine-tuned to ctx_len 4096.

  • Likely better. Please test.

RWKV-4-Pile-14B-20230213-8019.pth : Trained on the Pile for 331B tokens.

  • Pile loss 1.7579
  • LAMBADA ppl 3.81, acc 71.05%
  • PIQA acc 77.42%
  • SC2016 acc 75.57%
  • Hellaswag acc_norm 70.24%
  • WinoGrande acc 62.98%

ctx4096-test282: use this when you need long ctxlen. might be slightly weaker for short ctxlens.