SFTTrainer runs supervised fine-tuning on a model and training dataset. For clinical extraction, examples are formatted as system, user, and assistant messages, then exposed through a text column for next-token learning. section 7.3.3
Configure the training run through SFTConfig:
- Set epochs, per-device batch size, learning rate, scheduler, warmup, precision, sequence length, and the dataset text field. section 7.3.3
- Pass the model, training dataset, and configuration to SFTTrainer, then call trainer.train(). section 7.3.3
Listing 7.11 Configuring and training QLoRA
from trl import SFTTrainer, SFTConfig #1
training_args = SFTConfig(
output_dir="./qlora_output",
num_train_epochs=3, #2
per_device_train_batch_size=8,
gradient_accumulation_steps=1,
learning_rate=2e-4, #3
lr_scheduler_type="cosine", #4
warmup_steps=7, #5
fp16=False,
bf16=True, #6
logging_steps=10,
save_strategy="no",
max_length=MAX_SEQ_LENGTH,
dataset_text_field="text", #7
report_to="none",
)
trainer = SFTTrainer(
model=model,
train_dataset=train_dataset,
args=training_args,
)
trainer.train()
See a complete QLoRA training configuration and SFTTrainer invocation. section 7.3.3
Clinical extraction
After fine-tuning, the model achieved 100% schema compliance on 40 test samples with a minimal prompt, compared with 85% for the baseline strict prompt. section 7.3.3


