/train-student-gpt

Training a small language model from scratch on student lecture commentary

Primary LanguagePythonMIT LicenseMIT

Train Student GPT

Description

Train a small Generative Pre-trained Transformer to generate student lecture commentary data from SIGHT (Wang et. al., 2023).

Getting Started

To train the model, you can run:

python run.py --mode train

To use the model, you can run:

python run.py --mode generate --prompt "### Lec 29 | MIT 18.01 Single Variable Calculus, Fall 2007" --max_new_tokens 300

Results

Examples from Training Data

Processed from SIGHT data in comments.json. The format is "### {Lecture title}\n{Student comment}.

### 4. Factorization into A = LU
Thank you for your leasons!

### Lec 2 | MIT 18.01 Single Variable Calculus, Fall 2007
I sure will pay it back hundredfold. Thanks!!!

### 2. Conditioning and Bayes' Rule
amazing explanations

Training

The results of training for 27 minutes on an NVIDIA A100-80GB:

Loss Curve

The generations include several comments and titles which appear realistic relative to the model size.

Examples from Generated Data

Prompt was "### Lec 29 | MIT 18.01 Single Variable Calculus, Fall 2007". The following are the best generated video titles and comments chosen from the model's output.

### Lec 1 | MIT 18.01 Single Variable Calculus, Fall 2007
He the best.

### Lec 24 | MIT 18.01 Single Variable Calculus, Fall 2007
Thanks

### 1. Introduction to Statistics
this course 😂😂

Credits:

Andrej Karpathy for model code https://www.youtube.com/watch?v=kCc8FmEb1nY

Wang et. al., 2023 for data: https://github.com/rosewang2008/sight/