About

This repo is for running inverse scaling examples. There is a colab set up for it, which you can find in the task spreadsheet.

Running on NYU

To run on NYU:

Installing

1. Follow these Getting Started instructions to get connected to Greene.
2. Follow these Singularity instructions up until Install packages with the following differences:
   1. Instead of cuda11.2-cudnn8-devel-ubuntu20.04.sif, use cuda11.3.0-cudnn8-devel-ubuntu20.04.sif
   2. Instead of overlay-7.5GB-300K.ext3.gz use overlay-10GB-400K.ext3
3. Activate the Singularity image with the overlay
   1. Remember to run source /ext3/env.sh (or whatever you called it when setting up the image) to activate the Python environment.
4. cd to /ext3 and run git clone https://github.com/naimenz/inverse-scaling-eval-pipeline to get a copy of the code.
5. Run pip install . to install the inverse-scaling-eval-pipeline package.
6. Run the command python -m pip install torch==1.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html to install the correct version of PyTorch.

Running

1. Copy the example.sbatch script included under the /ext3/inverse-scaling-eval-pipeline/scripts directory to somewhere outside the image, e.g. your home or scratch.
2. There are two options for pointing to your data:
   1. Put your data in /ext3/inverse-scaling-eval-pipeline/data and use the option --data as in the script.
   2. Put your data elsewhere and use the option --dataset-path to point to it.
3. For --exp-dir, give the absolute path of the directory you want the results to be saved in.
4. Remember to add the flag --use-gpu only for HuggingFace models (GPT-2, GPT-Neo) and to add the flag --batch-size n (with n > 1) only for OpenAI API models (GPT-3)
5. Submit your .sbatch file as a job with sbatch example.sbatch
6. Run the plotting file by activating the Singularity image and running python /ext3/inverse-scaling-eval-pipeline/eval_pipeline/plot_loss.py </path/to/results/dir>

---

Let me know which parts of these instructions are incorrect/unclear!