https://research.ibm.com/blog/retrieval-augmented-generation-RAG
The project goal is to create an RAG model that can have highly flexible domain knowledge.
All intensive computations are done on MSOE's HPC cluster "ROSIE" https://www.msoe.edu/about-msoe/news/details/meet-rosie/
Any non-intensive computations are done on my at-home Unraid server with an Nvidia Tesla T4
https://yangkky.github.io/2019/07/08/distributed-pytorch-tutorial.html
- Understand transformers
- Crate a transformer network that can be trained across multiple GPUs!
- Read "Attention is all you need"
- Understand GPT models
- Develop a way to replicate the "chat" experience that comes with chat-gpt
- Use a "semantic search engine" to find files that contain info related to a search (similar to google's page rank system)
- Combine a fine tuned GPT model to a specific data set with a semantic search engine
- Test the results