Step 6: Configure Training

from transformers import TrainingArgumentsfrom trl import SFTTrainer training_args = TrainingArguments(    output_dir="outputs/ch2_domain_mistral",    num_train_epochs=3,    per_device_train_batch_size=2,    gradient_accumulation_steps=8,    learning_rate=2e-4,    warmup_ratio=0.03,    logging_steps=20,    save_strategy="epoch",    fp16=True,    report_to="none") trainer = SFTTrainer(    model=model,    train_dataset=dataset,    dataset_text_field="text",    tokenizer=tokenizer,    max_seq_length=1024,    packing=True,    args=training_args)

COPY

Code Breakdown

• What happens in this step
• You define the training settings (TrainingArguments).
• You create an SFTTrainer that knows how to fine-tune your LoRA-wrapped model on your dataset.
• Training does not start until Step 7 (trainer.train()).
• TrainingArguments: the key knobs
• per_device_train_batch_size=2 sets the batch size per GPU.
• gradient_accumulation_steps=8 simulates a larger batch by accumulating gradients. Effective batch size is about 2 × 8 = 16.
• learning_rate=2e-4 is a common starting point for LoRA/QLoRA adapter training.
• warmup_ratio=0.03 warms up the learning rate at the beginning for stability.
• fp16=True enables mixed precision to reduce memory and speed up training.
• save_strategy="epoch" saves a checkpoint after each epoch.
• output_dir=... is where outputs and checkpoints go.
• report_to="none" keeps the run simple (no external tracking).
• SFTTrainer: connecting model + data
• model=model should be the model after Step 4 (LoRA adapters attached).
• train_dataset=dataset is what you loaded in Step 5.
• dataset_text_field="text" tells the trainer which column contains the prompt string.
• max_seq_length=1024 truncates or pads sequences to this max length.
• packing=True packs multiple short examples into one sequence for better GPU utilization.

At this point, everything is configured. Next, Step 7 runs the actual training loop.