Upload training_config.yml with huggingface_hub

training_config.yml

@@ -6,7 +6,7 @@ model:
   apply_lora_to_mlp: false
   apply_lora_to_output: false
   lora_rank: 32
-  lora_alpha: 64
+  lora_alpha: 32
   perception_tokens: 2
   use_clip: false
 tokenizer:
@@ -46,7 +46,7 @@ batch_size: 4
 optimizer:
   _component_: torch.optim.AdamW
   weight_decay: 0.001
-  lr: 1e-6
+  lr: 1e-7
   betas:
   - 0.9
   - 0.998
@@ -59,7 +59,7 @@ grad_clip:
   _component_: torch.nn.utils.clip_grad_norm
   max_norm: 2.0
   norm_type: 2
-epochs: 12
+epochs: 8
 max_steps_per_epoch: null
 gradient_accumulation_steps: 32
 compile: true