Update training script for ThaiLLM-8B-SFT-IQ

README.md

@@ -75,6 +75,63 @@ The model is fine-tuned from **ThaiLLM-8B-SFT** using supervised fine-tuning (SF
 | Batch size | 8 |
 
 Training data consists of Thai medical question–answer pairs with context grounding and citation supervision.
+- **Training Script**
+- ```bash
+  model_base, tokenizer = FastLanguageModel.from_pretrained(
+        base_model,
+        max_seq_length=max_seq,
+        load_in_4bit=True,
+        load_in_8bit=False,
+        full_finetuning=False,
+        device_map = "balanced",
+    )
+
+   model = FastLanguageModel.get_peft_model(
+        model_base,
+        r=r,
+        lora_alpha=r, # Set lora_alpha = r
+        lora_dropout=0,
+        bias="none",
+        target_modules=[
+            "q_proj", "k_proj", "v_proj", "o_proj",
+            "gate_proj", "up_proj", "down_proj"
+        ],
+        use_gradient_checkpointing="unsloth",
+        random_state=seed,
+        use_rslora=False,
+        loftq_config=None,
+    )
+  
+  trainer = SFTTrainer(
+        model=model,
+        tokenizer=tokenizer,
+        train_dataset=train_ds,
+        dataset_num_proc=4,
+        args=SFTConfig(
+            dataset_text_field="text",
+            per_device_train_batch_size=2,
+            gradient_accumulation_steps=8,
+            gradient_checkpointing = True,
+            warmup_ratio=0.1,
+            warmup_steps=5,
+            # max_steps = 1000,          # If max_steps is set → ignore num_train_epochs
+            num_train_epochs=3,     # Epoch ..
+            learning_rate=args.learning_rate,
+            logging_steps=1,
+            optim="adamw_8bit",
+            weight_decay=0.01,
+            lr_scheduler_type="cosine",
+            fp16=not torch.cuda.is_bf16_supported(),
+            bf16=torch.cuda.is_bf16_supported(),
+            seed=args.seed,
+            report_to=["tensorboard"],
+            output_dir=f"{paths['log_path']}",
+            logging_dir=f"{paths['log_path']}",
+        ),
+    )
+
+     trainer.train()
+  ```
 
 ---