What is the loss Function when fine-tuning LlamaV2

Models
1

When fine-tuning LlamavaV2, what is the loss function that is being used? For instance, here’s the training snippet from Llama Recipes Github fine-tuning script. What is the loss function here?

from transformers import TrainerCallback
from contextlib import nullcontext
enable_profiler = False
output_dir = "tmp/llama-output"

config = {
    'lora_config': lora_config,
    'learning_rate': 1e-4,
    'num_train_epochs': 1,
    'gradient_accumulation_steps': 2,
    'per_device_train_batch_size': 2,
    'gradient_checkpointing': False,
}
from transformers import default_data_collator, Trainer, TrainingArguments
# Define training args
training_args = TrainingArguments(
    output_dir=output_dir,
    overwrite_output_dir=True,
    bf16=True,  # Use BF16 if available
    # logging strategies
    logging_dir=f"{output_dir}/logs",
    logging_strategy="steps",
    logging_steps=10,
    save_strategy="no",
    optim="adamw_torch_fused",
    max_steps=total_steps if enable_profiler else -1,
    **{k:v for k,v in config.items() if k != 'lora_config'}
)
with profiler:
    # Create Trainer instance
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        data_collator=default_data_collator,
        callbacks=[profiler_callback] if enable_profiler else [],
    )
    # Start training
    trainer.train()
1 Like