Fine-Tuning with LoRA

Source: Unsloth Documentation

Model Loading with 4-bit Quantization

from unsloth import FastLanguageModel

# Load model with 4-bit quantization
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/llama-3-8b-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,  # QLoRA
    dtype=None,
)

Apply LoRA Adapters

# Apply LoRA adapters
model = FastLanguageModel.get_peft_model(
    model,
    r=16,  # Rank: 8, 16, 32, 64, 128
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    lora_alpha=16,  # Equal to r, or 2*r for aggressive learning
    lora_dropout=0,  # 0 is optimized
    bias="none",  # "none" is optimized
    use_gradient_checkpointing="unsloth",  # 30% memory reduction
    random_state=3407,
    use_rslora=False,
    loftq_config=None,
)

Training Configuration

from transformers import TrainingArguments
from trl import SFTTrainer

training_args = TrainingArguments(
    output_dir="./outputs",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    warmup_steps=5,
    max_steps=60,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=1,
    optim="adamw_8bit",
    weight_decay=0.01,
    lr_scheduler_type="linear",
)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=2048,
    args=training_args,
)