Large Language Models for Efficient Data Annotation and Model Fine-Tuning with Iterative Active Learning

This framework combines human expertise with the efficiency of Large Language Models (LLMs) like OpenAI's GPT-3.5 to simplify dataset annotation and model improvement. The iterative approach ensures the continuous improvement of data quality, and consequently, the performance of models fine-tuned using this data. This not only saves time but also enables the creation of customized LLMs that leverage both human annotators and LLM-based precision.

Architecture

Features

1. Dataset Uploading and Annotation

   • Upload CSV datasets.
   • Leverage GPT-3.5 to automatically annotate datasets.
   • Preview the annotations, highlighting low-confidence score rows.

2. Manual Annotation Corrections

   • Display the annotated dataset for user-based corrections.
   • User can update labels for specific rows.

3. CleanLab: Confident Learning Approach

   • Utilizes confident learning to identify and rectify label issues.
   • Automatically displays rows with potential label errors for user-based corrections.

4. Data Versioning and Saving

   • Merge user corrections with the annotated dataset.
   • Advanced data versioning ensures unique dataset versions are saved for every update.

5. Model Training

   • Train a BERT model on the cleaned dataset.
   • Track and reproduce model versions seamlessly using MLflow.

Setup

Prerequisites

1. Install the required packages:

   pip install -r requirements.txt
   

Running the Tool

1. Start the FastAPI backend:

   uvicorn app:app --reload
   

2. Run the Streamlit app:

   streamlit run frontend.py
   

3. Launch MLflow UI: To view models, metrics, and registered models, you can access the MLflow UI with the following command:

   mlflow ui
   

4. Access the provided links in your web browser:

   • For the main application, access the Streamlit link.
   • For MLflow's tracking interface, by default, you can navigate to http://127.0.0.1:5000.

5. Follow the on-screen prompts to upload, annotate, correct, and train on your dataset.

Why Confident Learning?

Confident learning has emerged as a groundbreaking technique in supervised learning and weak-supervision. It aims at characterizing label noise, finding label errors, and learning efficiently with noisy labels. By pruning noisy data and ranking examples to train with confidence, this method ensures a clean and reliable dataset, enhancing the overall model performance.

License

This project is open-sourced under the MIT License.