diogodsa/clevergpt

Training GPTs to solve interaction nets

CleverGPT

Inspired by Victor Taelin's tweet about GPTs failing to be able to perform straightforward step-by-step reduction for interaction nets, this repo finetunes gpt-3.5-turbo with synthetic data to correctly perform step-by-step reduction for interaction nets.

How it works

Interaction nets similar to Taelin's tweet are pretty easy to generate: choose a token length, and then randomly pick tokens up to the length. Here we randomly choose up to 21 tokens; since that results in over 5 trillion possible interaction nets, training on a small subset of the nets should put to rest the idea that the model is just memorizing the output: if it's correctly solving a large percentage of the nets, it must have learned to generally apply an underlying pattern.

We train on 2000 randomly generated nets, and use a validation set also of 2000. 2k samples appears to be good enough to get very close to perfect performance at any size in the training range; originally we trained on 200 samples, which provided near-perfect performance on <10 token nets but could sometimes struggle with ones approaching 21 tokens. 2000 samples is still far below the 5 trillion valid 21-token inputs, so the model isn't memorizing the outputs, it's just getting better at applying the pattern it's learned.

Rather than using the #A A# #B B# notation Taelin originally used, we just use A B C D as tokens, to avoid possible model tokenization issues.

Running the code

To get started:

npm install # install deps
npx tsc # compile typescript

Make sure to add a .env file with the following env vars:

OPENAI_API_KEY="your-key-here"
MODEL_NAME="eventually once you have a trained model, put the model ID here"

To train a model, run:

node build/index.js

To try a random interaction net, update your .env file with the model ID that you trained and run:

node build/test.js