Tweet-Auth-LoRADistil is a machine learning solution that identifies disaster-related tweets in real-time. This repository leverages a fine-tuned DistilBERT model, modified with LoRA (Low-Rank Adaptation) and optimized with 8-bit quantization for efficient use, even in constrained environments. With this setup, users can discern between disaster-related and non-disaster tweets, making it a valuable tool for emergency response, social media monitoring, and disaster tracking applications.
Tweet-Auth-LoRADistil offers a balance of speed, accuracy, and memory efficiency by fine-tuning a light language model (DistilBERT) specifically on disaster detection tweets and enhancing it with LoRA layers. By using Docker, this model is easily accessible and deployable across various platforms, enabling real-time insights with minimal setup.
- Model Quantization: The model is quantized to 8-bit integers, which reduces its size but may affect accuracy marginally.
- LoRA Usage: LoRA layers are applied to key components of DistilBERT, allowing fine-tuning without large memory requirements.
- Docker Deployment: The project is available as a Docker image (
Tweet-Auth-LoRADistil) on Docker Hub, ensuring compatibility and ease of deployment.
- Disaster Response: Automatically classify incoming social media feeds to identify urgent disaster-related information.
- Social Media Monitoring: Filter for relevant tweets in disaster-prone areas to assess real-time situations.
- Media and Journalism: Track and verify incidents reported on social media to assist in news verification.
- DistilBERT: DistilBERT is a compact version of the BERT model, providing around 60% of the original BERT performance with just half its parameters. It’s suitable for tasks requiring both performance and efficiency.
- LoRA: LoRA (Low-Rank Adaptation) modifies large transformer models by introducing trainable rank-decomposed matrices in place of some of the weight matrices, making fine-tuning on specific tasks efficient and resource-friendly.
DistilBERT is a smaller, faster version of BERT with comparable performance. It is ideal for situations that need high-quality text representations while minimizing memory usage and computational costs, such as real-time tweet analysis.
- Efficient Memory Usage: With LoRA and 8-bit quantization, the model is efficient on memory.
- Real-Time Inference: Flask-based UI allows for quick, user-friendly interaction.
- Portable: Available as a Docker image for simplified deployment.
- Quantization Trade-offs: 8-bit quantization may slightly reduce model accuracy.
- Limited Context Understanding: DistilBERT’s smaller size could mean it misses some context nuances compared to larger BERT-based models.
- LoRA-enhanced DistilBERT: Fine-tuned on disaster tweets dataset.
- Flask UI: User-friendly interface for tweet classification.
- Dockerized: Available as a Docker image for cross-platform compatibility.
- Quantization: Model quantized for reduced size and memory usage.
Here’s the flowchart showing the overall structure of the process, from data ingestion to Docker deployment.
To get started with Tweet-Auth-LoRADistil using Docker, follow these steps:
-
Pull the Docker Image:
docker pull munjpatel2003/tweet-auth-lora-distil:latest
-
Run the Docker Container:
docker run -p 7860:7860 munjpatel2003/tweet-auth-lora-distil:latest
-
Access the Web Interface: Open your web browser and navigate to
http://localhost:5000to start analyzing tweets.
- LoRA Layers: Only fine-tuned layers use LoRA, reducing memory overhead and speeding up training.
- Dynamic Quantization: Compressing the model to 8-bit integer weights reduces memory usage with minimal performance drop.
- DistilBERT Base: Chosen for its balance between processing efficiency and quality in NLP tasks.
This project is licensed under the MIT License - see the LICENSE file for details.