File size: 1,062 Bytes
39cbe88 b191360 0e8912f b191360 d2f3da9 b191360 39cbe88 b191360 d8d732b b191360 d2f3da9 6d74069 d2f3da9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 |
---
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%
|