Model fine-tuning with Hugging Face

Fine-Tuning with Llama 3

Francesca Donadoni

Curriculum Manager, DataCamp

What do we need to conduct fine-tuning?

Language model + tokenizer (a LLama model, such as TinyLLama-v0)
Training dataset (the Bitext customer service dataset)
Training arguments
Conduct fine-tuning (SFTTrainer from TRL)
Evaluation benchmark or dataset

The training loop. Training needs training arguments, data, and the model, and produces a fine-tuned model. Evaluation occurs with the fine-tuned model and an evaluation dataset.

How to load models and tokenizers with Auto classes

model_name="Maykeye/TinyLLama-v0"

model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

tokenizer.pad_token = tokenizer.eos_token

¹ https://huggingface.co/docs/transformers/main/en/model_doc/auto

Defining training parameters with TrainingArguments

training_arguments = TrainingArguments(

    per_device_train_batch_size=1,

    learning_rate=2e-3,

    max_grad_norm=0.3,

    max_steps=200,

    ...
    gradient_accumulation_steps=2,
    save_steps=10,

)

¹ https://huggingface.co/docs/transformers/v4.40.1/en/main_classes/trainer#transformers.TrainingArguments

How to set up training with SFTTrainer

trainer = SFTTrainer(

    model=model,
    tokenizer=tokenizer,

    train_dataset=dataset,
    dataset_text_field='conversation',

    max_seq_length=250,

    args=training_arguments

)

Understanding fine-tuning results with SFTTrainer

trainer.train()

TrainOutput(global_step=200, training_loss=1.9401231002807617,
            metrics={'train_runtime': 142.5501, 
                     'train_samples_per_second': 2.806,
                     'train_steps_per_second': 1.403, 
                     'total_flos': 1461265827840.0, 
                     'train_loss': 1.9401231002807617, 
                     'epoch': 2.0})

How to evaluate a trained model Using ROUGE-1

ROUGE-1: Ratio of word overlap between a reference and generated text

import evaluate

rouge = evaluate.load('rouge')

predictions = ["hello there", "general kenobi"]
references = ["hello there", "master yoda"]

results = rouge.compute(predictions=predictions, references=references)
print(results)

{'rouge1': 0.5, 'rouge2': 0.5, 'rougeL': 0.5, 'rougeLsum': 0.5}

¹ https://huggingface.co/spaces/evaluate-metric/rouge

How to use the ROUGE-1 score

Use the evaluation set in evaluation_dataset

def generate_predictions_and_reference(dataset):
    predictions = []
    references = []
    for row in dataset:
        inputs = tokenizer.encode(row["instruction"], return_tensors="pt")

        outputs = model.generate(inputs)

        decoded_outputs = tokenizer.decode(outputs[0, inputs.shape[1]:], skip_special_tokens = True)

        references += [row["response"]]
        predictions += [decoded_outputs]
    return references, predictions

How to run ROUGE-1 on an evaluation set

references, predictions = generate_predictions_and_reference(evaluation_dataset)

rouge = evaluate.load('rouge')
results = rouge.compute(predictions=predictions, references=references)

print(results)

Finetuning vs no finetuning

Fine-tuned

{'rouge1': 0.22425812699023645,
 'rouge2': 0.039502543246449,
 'rougeL': 0.1501513006868983,
 'rougeLsum': 0.18685597710721613}

No fine-tuning

{'rouge1': 0.1310928764315105,
 'rouge2': 0.04581654122835097,
 'rougeL': 0.08415351421221628,
 'rougeLsum': 0.1224749866097021}

Let's practice!

Fine-Tuning with Llama 3