Eric Wang 3 роки тому
батько
коміт
b948f892ba
1 змінених файлів з 2 додано та 2 видалено
  1. 2 2
      finetune.py

+ 2 - 2
finetune.py

@@ -31,7 +31,7 @@ def train(
     micro_batch_size: int = 4,
     num_epochs: int = 3,
     learning_rate: float = 3e-4,
-    cutoff_len: int = 512,
+    cutoff_len: int = 256,
     val_set_size: int = 2000,
     # lora hyperparams
     lora_r: int = 8,
@@ -43,7 +43,7 @@ def train(
     ],
     # llm hyperparams
     train_on_inputs: bool = True,  # if False, masks out inputs in loss
-    group_by_length: bool = True,  # faster, but produces an odd training loss curve
+    group_by_length: bool = False,  # faster, but produces an odd training loss curve
 ):
     print(
         f"Training Alpaca-LoRA model with params:\n"