ML+Coffee: LoRA Finetuning

We're still looking for volunteers for our November 11 ML+Coffee! We have 30-60 minute slots available for anyone interested in sharing a project, demoing a tool or method, or discussing a paper. No formal presentation needed. It's a supportive space for researchers and practitioners across campus to share ideas and get feedback.

To fill an hour of our schedule, I'll lead a discussion and demo of LoRA finetuning of LLMs/VLMs for research applications. General-purpose AI models don't always perform well on the specialized datasets researchers work with. LoRA offers an inexpensive and powerful way to adapt existing models to your own data and tasks without training from scratch.

We'll use an exercise from this year's Machine Learning Marathon, finetuning a 3B vision-language model to transcribe historical German handwriting. You can find the exercise, code, and setup instructions in the new ML+X Nexus guide to LoRA finetuning.

I'll demonstrate the AWS workflow, covering how LoRA reduces GPU memory requirements (and spending), how to estimate and manage cloud training costs (about $0.75 for this exercise), and how to evaluate improvements in model accuracy. Evan Linck, ML Facilitator at CHTC, will join us to discuss how researchers can run similar workflows for free on CHTC.

November 11 ML+Coffee

• When: Wednesday, 9-11am CT
• Where: Room 1145, Discovery Building (330 N. Orchard St.)
• RSVP: https://forms.gle/ujz6hVTq9WNcjBct9
• Volunteer to lead a discussion: https://forms.gle/F5LYqnY5TMiZrzXW6
   

Coffee and pastries provided ☕, courtesy of our sponsors.