CS336 Spring 2026 Assignment 4: Data

For a full description of the assignment, see the assignment handout at cs336_assignment4_data.pdf

If you see any issues with the assignment handout or code, please feel free to raise a GitHub issue or open a pull request with a fix.

Setup

This directory is organized as follows:

• ./cs336_basics: This module contains the staff implementation of the language model from assignment 1. You will use this training code to train an LM on your filtered data. You should not modify the training logic, since your leaderboard submission must use it exactly.
• ./cs336_data: This folder is basically empty! This is the module where you will implement code to filter and process the data.

Visually, it should look something like:

.
├── cs336_basics  # A python module named cs336_basics
│   └── ... an optimized training implementation ...
├── cs336_data  # TODO(you): code that you'll write for assignment 4
│   ├── __init__.py
│   └── ... TODO(you): any other files or folders you need for assignment 4 ...
├── README.md
├── pyproject.toml
└── ... TODO(you): other files or folders you need for assignment 4 ...

As in previous assignments, we use uv to manage dependencies.

Downloading data

For students

Data is available at /shared-data, such as uv run modal shell ./scripts/download_data.py::main --cmd "ls -l /shared-data"

To only download the files needed for running offline, run uv run scripts/download_data.py --offline-only. To download all data for the students, run uv run modal run scripts/download_data.py

For non-students

To only download the files needed for running offline, run uv run scripts/download_data.py --offline-only.

Implement the method is_english in ./cs336_data/wet_files.py before downloading the full non-offline data.

Modal

Remove environment_name from shared_data_volume in ./cs336_data/modal_utils.py

uv run modal run scripts/download_data.py

Non-modal

Change the path in ./cs336_data/common.py and run uv run scripts/download_data.py.

Consider changing n_files for EnglishWetFiles in ./cs336_data/wet_files.py if you want to download less than 2.5k WET files.

Training on Modal

The Modal entrypoint in ./scripts/train.py contains the full training config; pass the path to your GPT-2-tokenized training data with --train-bin.

The final training run uses 8 B200 GPUs:

uv run modal run scripts/train.py --train-bin /root/data/your_data.bin

Submitting

To submit, run ./test_and_make_submission.sh . This script will install your code's dependencies, run tests, and create a zip with the output. We should be able to unzip your submitted tarball and run ./test_and_make_submission.sh to verify your test results.