Project 3: Text Models

two cartoon llamas. one is writing on a piece of paper with a cup of coffee to the side and the other is reading a piece of paper.

Build a local, private LLM system using Ollama and RAG (Retrieval-Augmented Generation) that runs entirely on your computer. You'll create a chatbot that answers questions based on documents you provide.

Alternatively you may train a microGPT instead (see bottom of this text for directions)

Questions to consider:

What knowledge do you have that deserves a dedicated AI assistant?
Who would benefit from easy access to this information?
What would a corporate LLM get wrong about this topic?
Why does privacy matter for this knowledge?
How might keeping AI local change what's possible?

---------------------------------------------------------

Create a curated document collection

see:

See https://github.com/AaratiAkkapeddi/llllm/blob/main/DATA_TIPS.md

Working with RAG:

See https://github.com/AaratiAkkapeddi/llllm

---------------------------------------------------------

INSPO

Computational Mama (aka Ambika Joshi), Sun Shines Bright: https://solararchive.cmama.xyz/

---------------------------------------------------------

MICROGPT

Alternatively, if you would rather train a model from scratch only on your own data you could work with Microgpt.

See https://github.com/AaratiAkkapeddi/microgpt for instructions

---------------------------------------------------------

Inspo:

Allison Parish

Apotropaic Variations

Each 200 (of 10,000) Apotropaic Variations is one of a set of 50, containing 200 computer-generated magic words, printed so that they can be easily cut out. This allows each word to be worn on one’s person or to be ingested for protective effect.
---------------------------------------------------------

**DUE WEEK 6***

In discord submit a google drive link containing:

your text corpus (5-20 documents you used)

---------------------------------------------------------

**DUE WEEK 7***

Train your model using the RAG method OR the Microgpt method we went over in class and your own text corpus.

In discord submit a google drive link containing:

In discord submit a google drive link containing:

your text corpus (5-20 documents you used)
question-answer pairs from your trained chatbot (can be screenshots of you interacting with it) OR output from your MicroGPT model
A text document answering:
- what did you use to teach your AI and why?
- What context could this be used (Art is a welcome context!)? Who is this for?

Back to projects