fix: use /tmp for all output directories

train.py

@@ -176,7 +176,7 @@ trainer = GRPOTrainer(
     processing_class=tokenizer,
     reward_funcs=[reward_fn],
     args=GRPOConfig(
-        output_dir="./deceit-1.5b",
+        output_dir="/tmp/deceit-1.5b",
         max_steps=150,
         per_device_train_batch_size=4,
         num_generations=4,
@@ -195,8 +195,8 @@ wandb.finish()
 print("Training done!")
 
 # Save Level 1 checkpoint
-model.save_pretrained("deceit-1.5b-l1")
-tokenizer.save_pretrained("deceit-1.5b-l1")
+model.save_pretrained("/tmp/deceit-1.5b-l1")
+tokenizer.save_pretrained("/tmp/deceit-1.5b-l1")
 print("Level 1 checkpoint saved locally")
 
 # Load Level 2 dataset
@@ -268,7 +268,7 @@ trainer_l2 = GRPOTrainer(
     processing_class=tokenizer,
     reward_funcs=[reward_fn_l2],
     args=GRPOConfig(
-        output_dir="./deceit-1.5b-l2",
+        output_dir="/tmp/deceit-1.5b-l2",
         max_steps=80,
         per_device_train_batch_size=4,
         num_generations=4,
@@ -287,8 +287,8 @@ wandb.finish()
 print("Level 2 training done!")
 
 # Save final model
-model.save_pretrained("deceit-1.5b-final")
-tokenizer.save_pretrained("deceit-1.5b-final")
+model.save_pretrained("/tmp/deceit-1.5b-final")
+tokenizer.save_pretrained("/tmp/deceit-1.5b-final")
 model.push_to_hub(HF_REPO_ID)
 tokenizer.push_to_hub(HF_REPO_ID)
 print(f"Final model saved to {HF_REPO_ID}")