Welcome to the Finetuned Chatbot repository! 🚀 This project showcases a chatbot application built using Streamlit, leveraging the powerful facebook/bart-base model fine-tuned on the Bitext-customer-support-llm-chatbot-training-dataset. The primary goal? To test the conditional generation capabilities of BART while delivering contextually relevant and well-structured responses. 💬✨
You can find the fine-tuned model and related files here on Google Drive. 💾✨ .
Here are the key metrics from the fine-tuning process:
- Runtime: 15,521.44 seconds
- Steps: 110
- Evaluation Loss: 0.1015
- Evaluation Runtime: 411.59 seconds
- Samples per Second: 13.059
- Steps per Second: 1.088
- Total FLOPs: 39,322,513,928,355,840
- Training Epochs: 3
- Global Steps: 5376
- Gradient Norm: 0.3138
- Learning Rate: 9.67e-8
- Training Loss: 0.116
- Overall Training Loss: 0.2455
- Training Runtime: 15,513.66 seconds
- Samples per Second: 4.157
- Steps per Second: 0.347
The chatbot is powered by the facebook/bart-base model, fine-tuned for this specific use case. You can explore the fine-tuning process in detail through this Notebook implementation. 📒✨
This repository features a basic Streamlit app that brings the chatbot to life! While it’s still a work in progress, here’s what it currently offers:
- ✅ Contextually relevant responses
- ✅ Well-structured output based on the fine-tuned dataset
Check out the demo in action:
We’ve got exciting plans for the next iterations:
- RAG (Retrieval-Augmented Generation): Integrate external knowledge sources for enhanced contextuality.
- Conversational Memory: Enable the chatbot to maintain conversation context across multiple exchanges.
This repository is licensed under the MIT License. 📝 Feel free to explore, use, and contribute!
This project was created as a test of BART’s conditional generation capabilities—and it delivered brilliantly! ✨ The model’s ability to generate structured responses opens up exciting possibilities for similar applications.
Stay tuned for more updates and enhancements!