This educational repository contains code for fine-tuning GPT-2 on dialogue from the TV series Spongebob Squarepants.

The code for fine-tuning GPT-2 comes from Karpathy's NanoGPT and uses PyTorch.

0. Setup
1. Scraping
2. preprocessing
3. training
4. sampling

conda env create -f environment.yml

Spongebob transcripts can be scraped from the fandom website using the scrape.py script and Selenium. The scripts for each episode are stored in data, and the full transcript of every episode is stored in spongebob_anthology.txt.

The preprocess.py script takes the full transcript and simply removes non-dialogue lines. The result is stored in spongebob_anthology_cleaned.txt.

After a clean version of the transcript is created, navigate to data/spongebob/ and run the prepare.py (from NanoGPT) script to make the training and test splits.

Use train.py to train the model, with the either the following arguments

• python train.py config/train_spongebob_char.py to train from scratch and output to out_spongebob_char.
• python train.py config/finetune.py to train from GPT2 and output to out_spongebob_gpt.

Use sample.py to sample from the model

Example:

• python sample.py --out_dur={specified output dir} --start="{starting text}" --num_samples={number of samples}