How can I fine-tune a pre-trained transformer model for sentiment analysis?
Asked on Oct 01, 2025
Answer
Fine-tuning a pre-trained transformer model for sentiment analysis involves adapting the model to a specific sentiment classification task using a labeled dataset. This process typically requires a smaller dataset and less computational power compared to training a model from scratch.
<!-- BEGIN COPY / PASTE -->
from transformers import BertTokenizer, BertForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset
# Load pre-trained model and tokenizer
model_name = "bert-base-uncased"
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForSequenceClassification.from_pretrained(model_name, num_labels=2)
# Load and preprocess dataset
dataset = load_dataset("imdb")
def tokenize_function(examples):
return tokenizer(examples["text"], padding="max_length", truncation=True)
tokenized_datasets = dataset.map(tokenize_function, batched=True)
# Define training arguments
training_args = TrainingArguments(
output_dir="./results",
evaluation_strategy="epoch",
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
num_train_epochs=3,
weight_decay=0.01,
)
# Initialize Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["test"],
)
# Fine-tune the model
trainer.train()
<!-- END COPY / PASTE -->Additional Comment:
- Start by selecting a pre-trained transformer model, such as BERT, which is well-suited for text classification tasks.
- Use a tokenizer to preprocess your text data, ensuring it matches the input format expected by the model.
- Load a sentiment analysis dataset, such as IMDb, and split it into training and evaluation sets.
- Set up training arguments, specifying parameters like batch size, number of epochs, and weight decay for regularization.
- Use the Trainer class from the Transformers library to handle the training loop and evaluation.
- Fine-tuning adjusts the model's weights to better predict sentiment labels based on your specific dataset.
Recommended Links: