Add code

README.md

@@ -25,3 +25,65 @@ https://huggingface.co/datasets/tisu1902/var-full-no-think
 
 ## Wandb
 https://wandb.ai/quangphamm1902/huggingface/runs/4bv2nr1q?nw=nwuserquangphamm1902
+
+```
+from trl import SFTTrainer, SFTConfig
+
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=train_dataset,
+    eval_dataset=val_dataset,
+    args=SFTConfig(
+        dataset_text_field="text",
+        max_seq_length=18432,
+        
+        # Reduce effective batch for more steps
+        per_device_train_batch_size=4,  # Down from 6
+        gradient_accumulation_steps=2,  # Down from 6
+        # Effective batch = 8 (instead of 36)
+        # New steps per epoch = 410 / 8 ≈ 51
+        # Total steps = 51 × 3 = 153 steps
+        per_device_eval_batch_size=1,
+        
+        # Training
+        num_train_epochs=3,
+        learning_rate=1e-4,
+        warmup_ratio=0.1,
+        lr_scheduler_type="cosine",
+        
+        # Optimization
+        optim="adamw_8bit",
+        weight_decay=0.01,
+        max_grad_norm=0.3,
+        
+        # Memory
+        fp16=not torch.cuda.is_bf16_supported(),
+        bf16=torch.cuda.is_bf16_supported(),
+        gradient_checkpointing=True,
+        
+        # Evaluation & Saving - ADJUSTED FOR FEWER STEPS
+        eval_strategy="steps",
+        eval_steps=25,
+        save_strategy="steps",
+        save_steps=25,
+        save_total_limit=3,
+        load_best_model_at_end=True,
+        metric_for_best_model="eval_loss",
+        greater_is_better=False,
+        
+        # Logging
+        logging_steps=5,
+        logging_first_step=True,
+        
+        # Other
+        seed=3407,
+        output_dir="outputs",
+        remove_unused_columns=False,
+        
+        # TRL specific
+        dataset_num_proc=4,
+        packing=False,
+    ),
+)
+```