|
--- |
|
language: |
|
- en |
|
tags: |
|
- pytorch |
|
- text-generation |
|
- causal-lm |
|
- rwkv |
|
license: apache-2.0 |
|
datasets: |
|
- the_pile |
|
|
|
--- |
|
|
|
# RWKV-4 430M |
|
|
|
## Model Description |
|
|
|
RWKV-4 430M is a L24-D1024 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 = 1024 |
|
|
|
Final checkpoint: |
|
RWKV-4-Pile-430M-20220808-8066.pth : Trained on the Pile for 333B tokens. |
|
* Pile loss 2.2621 |
|
* LAMBADA ppl 13.04, acc 45.16% |
|
* PIQA acc 67.52% |
|
* SC2016 acc 63.87% |
|
* Hellaswag acc_norm 40.90% |
|
|
|
## Warning: 4 / 4a / 4b models ARE NOT compatible!!! Use RWKV-4 unless you know what you are doing. |
|
|
|
With tiny attention (--tiny_att_dim 512 --tiny_att_layer 18): |
|
RWKV-4a-Pile-433M-20221223-8039.pth |
|
* Pile loss 2.2394 |
|
* LAMBADA ppl 10.54, acc 50.20% |
|
* PIQA acc 68.12% |
|
* SC2016 acc 63.55% |
|
* Hellaswag acc_norm 40.82% |
|
|
