try lower lr

allenai · Oct 8, 2024 · ad10add · ad10add
1 parent 230c8a9
commit ad10add
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/pdelfin/train/config/qwen2vl-7b-lora.yaml b/pdelfin/train/config/qwen2vl-7b-lora.yaml
@@ -54,8 +54,8 @@ hparams:
   gradient_accumulation_steps: 4
   gradient_checkpointing: false
   clip_grad_norm: 1.0
-  learning_rate: 3e-4
-  max_steps: 9000
+  learning_rate: 1e-4
+  max_steps: 20000
   pad_multiple_of: 16
   log_every_steps: 10
   eval_every_steps: 100