reducing example config max training token length for something more sane

Author: benjaminye

config.yml

@@ -61,7 +61,7 @@ training:
     warmup_ratio: 0.03
     lr_scheduler_type: "constant"
   sft_args:
-    max_seq_length: 5000
+    max_seq_length: 1024
     # neftune_noise_alpha: None
 
 inference:
@@ -80,4 +80,4 @@ qa:
     - verb_percent
     - adjective_percent
     - noun_percent
-    - summary_length
+    - summary_length