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%