An end-to-end LLM course that archived itself before graduation

What it does

This repository is a complete, free course—videos, articles, and code—for building a real-time financial-advisor LLM system. It walks through three pipelines: training an open-source model with QLoRA on a custom dataset; streaming financial news into Qdrant vector embeddings via Bytewax; and serving RAG-augmented inference through a Gradio UI using LangChain. The stack is glued together with Comet ML for tracking, Beam for serverless GPU compute, and AWS EC2 for streaming deployment.

The interesting bit

Rather than curating data by hand, the authors distilled GPT-3.5 into a financial Q&A dataset to specialize a smaller Falcon 7B model. The whole system is framed as a production-grade, three-pipeline architecture deployed across serverless and cloud infrastructure instead of a single local notebook.

Key highlights

• Fine-tunes an LLM with QLoRA using a GPT-3.5-distilled financial Q&A dataset.
• Streams and embeds live Alpaca financial news into Qdrant using Bytewax.
• Inference chain uses LangChain to merge vector retrieval, fine-tuned generation, and chat history.
• Training and inference run on Beam serverless GPUs; the streaming pipeline operates on AWS EC2 managed through GitHub Actions CI/CD.
• Course materials include YouTube videos and MIT-licensed code, though some articles live behind Medium’s paywall.

Caveats

• The authors archived the course because GenAI moves too fast for educational content to keep up; they now recommend a newer LLM Twin course.
• Reproducing the system requires juggling several freemium third-party accounts—Beam, Qdrant, Comet ML, Alpaca, and AWS.
• Local execution demands a CUDA GPU with up to 10 GiB of VRAM for training, though the README steers you toward cloud deployments.

Verdict

Worth browsing if you want a guided, opinionated map of how modern LLM pieces fit together in a real-ish system. Avoid if you need a maintained, current template; the authors have already moved on.