BlinkDL commited on
Commit
29fbd06
1 Parent(s): 5ef5c9f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +8 -6
README.md CHANGED
@@ -26,9 +26,11 @@ ctx_len = 1024
26
  n_layer = 32
27
  n_embd = 4096
28
 
29
- Preview checkpoint: RWKV-4-Pile-7B-20221030-6224.pth : Trained on the Pile for 257B tokens.
30
- * Pile loss 1.8553
31
- * LAMBADA ppl 4.36, acc 67.42%
32
- * PIQA acc 75.68%
33
- * SC2016 acc 72.58%
34
- * Hellaswag acc_norm 64.87%
 
 
26
  n_layer = 32
27
  n_embd = 4096
28
 
29
+ Final checkpoint: RWKV-4-Pile-7B-20221115-8047.pth : Trained on the Pile for 332B tokens.
30
+ * Pile loss 1.8415
31
+ * LAMBADA ppl 4.38, acc 67.18%
32
+ * PIQA acc 76.06%
33
+ * SC2016 acc 73.44%
34
+ * Hellaswag acc_norm 65.51%
35
+
36
+ I am fine-tuning it to ctx_len 4096.