This project implements a sentiment analysis model for Amazon product reviews using a fine-tuned DistilBERT model with LoRA (Low-Rank Adaptation). The model classifies reviews into three categories: positive, neutral, and negative.
- Load and preprocess data from a CSV file.
- Balance the dataset using resampling techniques.
- Train a DistilBERT model with cross-validation.
- Evaluate model performance using metrics like accuracy, precision, recall, and F1 score.
- Generate confusion matrices for visualizing predictions.
Make sure you have the following Python packages installed:
numpypandasmatplotlibseaborntorchtransformersdatasetsscikit-learnpeft
You can install the necessary packages using pip:
pip install numpy pandas matplotlib seaborn torch transformers datasets scikit-learn peftClone this repository:
git clone <repository-url>
cd <repository-folder>
Install the required packages as mentioned above.
Prepare your dataset and ensure it is in the correct format.
Data Preparation Place your dataset CSV file in the specified path (FILE_PATH in the code). Modify the COLUMNS_TO_KEEP constant if your dataset has different column names.
To train the model, run the following command:
python train.py
This will load the dataset, balance it, and start the training process. The trained model will be saved in the ./saved_model directory.
To test the model, you can use the following command:
python test.py
Make sure to set the TEST_FILE_PATH variable to point to your test data CSV file.
The training script will output average metrics (accuracy, precision, recall, F1 score) and display a confusion matrix of the predictions.
This project is licensed under the MIT License - see the LICENSE file for details.
Make sure to replace `<repository-url>` and `<repository-folder>` with the actual URL of your repository and the folder name, respectively. You can also adjust any sections to better match the specifics of your project or your coding style!