Project Chiron

Mission to see true active teaching made available to all students.

Previous Experiments

I've collected 350 year 11 maths mock exams (UK based) and conducted a series of investigations to better understand how large language models can be utilised in an education setting. Please see the dataset page for the details on the data.

BERT-based classifiers

The end model here simply consists of a pretrained transformer model (which produces a layer of embeddings) with a multilayer perceptron classifier head. The classifier head is trained on the specific maths problem.

Sentence embeddings

A multitude of different pretrained transformers have been used to extract sentence level embeddings from the processed student problem attempts. Please see the embeddings page for more info on these models.

classifier head

GPT-based inference

fine-tuning base GPT-3.0

zero/few-shot prompting GPT-3.5

Future Plan