Home

Welcome to the LLMHumanModeling wiki.

A brief abstract of this project can be found below as it was presented at the BYU CPMS Student Research Conference, Feb 25, 2023:

Large language models (LLMs) have historically been trained on massive sets of corpora and optionally fine-tuned to improve at a given task. This allows for coverage and diversity that human learning isn’t capable of; but versatility is not the only metric of LLM success. Many researchers are turning to LLMs as a resource for hypothesis testing and data synthesis, especially in research irreplicable in a human population due to ethical or temporal constraints. An LLM that better models human opinion and behavior would greatly aid such research; this is our model’s aim. To do so, we propose a fine-tuning process that imitates agentive characteristics of human learning such as curiosity, attention span, predisposition to contradiction, zone of proximal development, and other phenomena as described by learning-adjacent fields. The model may also include environmental influences such as socio-historic and algorithmic influences. It is our hope that this model architecture will enable to faster, safer and more representative ways to study human learning processes and conduct social science research.

This research draws on theories and practices of psychology, cognitive modeling, instructional design, computer science, and linguistics. Some foundational ideas include zone of proximal development (Vigotsky, 1978) and curricculum learning (Bengio et. al., 2009).

Helping out? Look at the following list of to-do's and corresponding deadlines:

• (3/23) Test search engine on wiki dataset

• (3/28) Choose/develop function to model curiosity (self-assesment curriculum learning?)

• (4/4) Build a model architecture (See Curriculum Learning: A Survey)

• Log learned files (percentage of file learned and order files were learned from)

• (4/8) First run model on lab computer (Formic?)

• (4/10) Second run model on lab computer (Formic?)

• (4/8) Find set of established linguistic evaluations

• (4/14) Evaluate perplexity

• (4/14) Evaluate training time and loss (against a normal model?)