훈련을 통한 파인튜닝

Python으로 배우는 LLM 입문

Jasmin Ludolf

Senior Data Science Content Developer, DataCamp

훈련 인수

from transformers import Trainer, 
TrainingArguments

training_args = TrainingArguments(

  output_dir="./finetuned",

  evaluation_strategy="epoch",

  num_train_epochs=3,

  learning_rate=2e-5,







)

TrainingArguments(): 훈련 설정 커스터마이즈
모든 매개변수는 공식 문서 참고
값은 용도, 데이터셋, 속도에 따라 다름
output_dir: 출력 디렉터리
eval_strategy: 평가 시점 — "epoch", "steps", "none"
num_train_epochs: 훈련 에포크 수
learning_rate: 옵티마이저용 학습률

훈련 인수

from transformers import Trainer, 
TrainingArguments

training_args = TrainingArguments(
  output_dir="./finetuned",
  evaluation_strategy="epoch",
  num_train_epochs=3,
  learning_rate=2e-5,

  per_device_train_batch_size=8,
  per_device_eval_batch_size=8,

  weight_decay=0.01,

)

per_device_train_batch_size 및 per_device_eval_batch_size: 배치 크기 정의
weight_decay: 과적합 방지를 위해 옵티마이저에 적용

Trainer 클래스

from transformers import Trainer, 
TrainingArguments

training_args = TrainingArguments(...)

trainer = Trainer(

    model=model,

    args=training_args,

    train_dataset=tokenized_training_data,

    eval_dataset=tokenized_test_data,

    tokenizer=tokenizer

)


trainer.train()

model: 파인튜닝할 모델
args: 훈련 인수
train_dataset: 훈련용 데이터
eval_dataset: 평가용 데이터
tokenizer: 토크나이저

훈련 반복 횟수: 데이터셋 크기, num_train_epochs, per_device_train_batch_size, per_device_eval_batch_size

Trainer 출력

{'eval_loss': 0.398524671792984, 'eval_runtime': 33.3145, 'eval_samples_per_second': 46.916, 
'eval_steps_per_second': 5.883, 'epoch': 1.0}
{'eval_loss': 0.1745782047510147, 'eval_runtime': 33.5202, 'eval_samples_per_second': 46.629, 
'eval_steps_per_second': 5.847, 'epoch': 2.0}
{'loss': 0.4272, 'grad_norm': 15.558795928955078, 'learning_rate': 2.993197278911565e-06, 
'epoch': 2.5510204081632653}
{'eval_loss': 0.12216147780418396, 'eval_runtime': 33.2238, 'eval_samples_per_second': 47.045, 
'eval_steps_per_second': 5.899, 'epoch': 3.0}
{'train_runtime': 673.0528, 'train_samples_per_second': 6.967, 'train_steps_per_second': 0.874, 
'train_loss': 0.40028538347101533, 'epoch': 3.0}
TrainOutput(global_step=588, training_loss=0.40028538347101533, metrics={'train_runtime': 673.0528, 
'train_samples_per_second': 6.967, 'train_steps_per_second': 0.874, 
'train_loss': 0.40028538347101533, 'epoch': 3.0})

파인튜닝된 모델 사용

new_data = ["This is movie was disappointing!", "This is the best movie ever!"]


new_input = tokenizer(new_data, return_tensors="pt", padding=True, truncation=True, max_length=64)


with torch.no_grad():
    outputs = model(**new_input)


predicted_labels = torch.argmax(outputs.logits, dim=1).tolist()

label_map = {0: "NEGATIVE", 1: "POSITIVE"}
for i, predicted_label in enumerate(predicted_labels):
    sentiment = label_map[predicted_label]
    print(f"\nInput Text {i + 1}: {new_data[i]}")
    print(f"Predicted Label: {sentiment}")

파인튜닝 결과

Input Text 1: This is movie was disappointing!
Predicted Sentiment: NEGATIVE

Input Text 2: This is the best movie ever!
Predicted Sentiment: POSITIVE

모델 및 토크나이저 저장

model.save_pretrained("my_finetuned_files")

tokenizer.save_pretrained("my_finetuned_files")

# Loading a saved model
model = AutoModelForSequenceClassification.from_pretrained("my_finetuned_files")
tokenizer = AutoTokenizer.from_pretrained("my_finetuned_files")

연습해 봅시다!

Python으로 배우는 LLM 입문