Skip to content

Repository files navigation

NRGPT

Code supplementary to accompany the NRGPT paper

Getting started

Initialize Environment

uv sync
source .venv/bin/activate
make data # takes a bit

All experiments use wandb by default. Create a .env file in the root directory with your wandb information:

WANDB_ENTITY=...

Experiments

Shakespeare Scaling (SSCL00)

sweeps/SSCL00_shakespeare_scaling.py

uv run python run_sweep_wandb.py --sweep ./sweeps/SSCL00_tst.py --num_agents 1 --gpus 0

Listops transitions (LOPS00)

sweeps/LOPS00_listops_training.py

uv run python run_sweep_wandb.py --sweep ./sweeps/LOPS00_listops_training.py --num_agents 1 --gpus 0

Listops deep train (LOPS01)

sweeps/LOPS01_listops_deep.py

uv run python run_sweep_wandb.py --sweep ./sweeps/LOPS01_listops_deep.py --num_agents 1 --gpus 0

OpenWebText

uv run python run_sweep_wandb.py --sweep ./sweeps/OWT00_owt_training.py --num_agents 1 --allgpus

About

Code accompanying experiments for our NRGPT paper

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages