Upload train.py with huggingface_hub

train.py

@@ -68,14 +68,15 @@ def main():
     # 7. Training Arguments (TRL v0.8.6 uses TrainingArguments from transformers)
     training_args = TrainingArguments(
         output_dir=OUTPUT_DIR,
-        per_device_train_batch_size=2,
-        gradient_accumulation_steps=4,
+        per_device_train_batch_size=1,
+        gradient_accumulation_steps=8,  # Compensate for smaller batch
         learning_rate=2e-4,
         logging_steps=10,
         num_train_epochs=1,
         optim="paged_adamw_32bit",
         fp16=True,
         group_by_length=True,
+        gradient_checkpointing=True,  # Save memory
         save_strategy="epoch",
         report_to="none",
         push_to_hub=True,
@@ -88,7 +89,7 @@ def main():
         train_dataset=dataset,
         peft_config=peft_config,
         formatting_func=formatting_prompts_func,
-        max_seq_length=2048,
+        max_seq_length=1024,  # Reduced for T4 memory
         tokenizer=tokenizer,
         args=training_args,
     )