BlinkDL commited on
Commit
5393d2f
1 Parent(s): dc1786f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +2 -2
README.md CHANGED
@@ -20,14 +20,14 @@ RWKV-4 3B is a L32-D2560 causal language model trained on the Pile. See https://
20
 
21
  Use https://github.com/BlinkDL/ChatRWKV to run it.
22
 
23
- New checkpoint: RWKV-4-Pile-3B-20221110-ctx4096.pth : Fine-tuned to ctx_len = 4096
24
  * LAMBADA ppl 5.25, acc 63.96%
25
  * PIQA acc 74.16%
26
  * SC2016 acc 70.71%
27
  * Hellaswag acc_norm 59.89%
28
  * ctx_len = 4096 n_layer = 32 n_embd = 2560
29
 
30
- Final checkpoint: RWKV-4-Pile-3B-20221008-8023.pth : Trained on the Pile for 331B tokens.
31
  * Pile loss 1.9469
32
  * LAMBADA ppl 5.24, acc 63.94%
33
  * PIQA acc 73.72%
 
20
 
21
  Use https://github.com/BlinkDL/ChatRWKV to run it.
22
 
23
+ Best checkpoint: RWKV-4-Pile-3B-20221110-ctx4096.pth : Fine-tuned to ctx_len = 4096
24
  * LAMBADA ppl 5.25, acc 63.96%
25
  * PIQA acc 74.16%
26
  * SC2016 acc 70.71%
27
  * Hellaswag acc_norm 59.89%
28
  * ctx_len = 4096 n_layer = 32 n_embd = 2560
29
 
30
+ Previous checkpoint: RWKV-4-Pile-3B-20221008-8023.pth : Trained on the Pile for 331B tokens.
31
  * Pile loss 1.9469
32
  * LAMBADA ppl 5.24, acc 63.94%
33
  * PIQA acc 73.72%