Teaching at Wharton has been incredibly rewarding, especially covering the latest advances in NLP. However, teaching and building an application are vastly different experiences. This contrast inspired me to embark on an NLP/LLM-based project that lets me explore the practical applications of the techniques I teach. Enter the Miata-bot—my Miata chatbot project, which I aim to develop over the next few months.
Why a Miata-bot?
I own a 2019 Miata, and as the saying goes, “Miata is always the answer” :)
While the plan is likely to evolve, here is the current outline:
-
Data Collection: Scrape and gather text data about Miatas from online conversations and forums. Fortunately, there are many active Miata communities that can provide valuable data.
-
Building an LLM Application: Use the collected data to fine-tune a language model, implementing techniques like fine-tuning and retrieval-augmented generation (RAG) to create a Miata-bot that can answer both common and uncommon questions about Miatas.
-
Evaluation: Generate question-answer pairs to evaluate the quality of the bot’s responses. This may involve some red teaming to test the robustness of the bot.
-
Tool Exploration: This project also serves as a way to explore the current ecosystem of packages such as LangChain, Hugging Face, Ray, etc., to build and deploy specialized LLMs.
-
Quantization: I am curious about quantizing the Miata-bot and evaluating its performance as a quantized model.
I am sharing the code for this project, which is available on GitHub. Currently, I am collecting data from the largest Miata forum.
While I don’t plan to publish all aspects of this work, I welcome collaboration. If you are a researcher or Miata enthusiast with a tech or AI/ML background and would like to contribute, please feel free to reach out.