update

examples/tutorials/dpo/ultrachat/step_2_train_sft_model.py

@@ -82,21 +82,19 @@ def main():
         name=args.dataset_name,
         split="train_sft",
         cache_dir=args.dataset_cache_dir,
-        # num_proc=args.num_workers if not args.dataset_streaming else None,
-        # streaming=True if args.dataset_streaming in ("true",) else False,
     )
 
     # 配置SFT参数 - 使用 TrainingArguments 而不是 SFTConfig
     training_args = TrainingArguments(
         output_dir=args.output_model_dir,
         num_train_epochs=1,
-        per_device_train_batch_size=args.batch_size,
-        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        per_device_train_batch_size=8,
+        gradient_accumulation_steps=4,
         save_strategy="steps",
         save_steps=500,
         save_total_limit=2,
         logging_steps=100,
-        learning_rate=args.learning_rate,
+        learning_rate=2e-5,
         warmup_ratio=0.03,
         lr_scheduler_type="cosine",
         bf16=torch.cuda.is_available(),