Bad Writing is "Fine": Tuning an LLM to Suggest Improvements

Description

Prototype an approach to fine-tune a large language model (LLM) to help diagnose areas to improve a specific writing product. For example, scientific papers require consistent language but in creative writing variety matters. Proposed steps are:

1. Writing Product: Coordinate with project mentors to choose a common and important writing product, such as a position paper or an academic conference. Identify/gather a rubric and a corpus.
2. Inject Bad Writing: For each element of the rubric, develop prompts for generative AI to decrease the quality of writing based on the rubric (i.e., make it worse). This will form a training data set of the good example and version worse on certain characteristics.
3. Fine Tune: Students will be expected to attempt to fine tune an LLM (e.g., LLAMA 2) based on this synthetically generated data
4. Evaluate: Research if tuning suggests better domain-specific areas to improve.

This project aligns with ongoing work with the USC Generative AI Center.

Students

• Anupam Patil

• Anuranjan Pandey

• Divyajyoti Panda

• Jize Luo

• Nilamadha Mohanty

• Rahul Tangsali

Advisors

• Benjamin Nye

What students will learn

Generative AI for large language models. Generating synthetic data for a rubric. Fine tuning a large language model, likely using CARC (the on campus computing cluster). Understanding intelligent tutoring system design fundamentals for modeling how experts diagnose issues from novices.