Marlowe Hackathon 2026

Materials for the DARC and Marlowe GPU hackathon at Stanford GSB, November 13, 2026.

Website: https://gsbdarc.github.io/marlowe-hackathon/

Participants fine-tune open models to answer three questions about CEO turnover in SEC 8-K Item 5.02 filings the way Jev, a large hosted model, answered them. They pick one of three tracks (Beginner, Intermediate, Advanced), start on the Yens, then scale the training to 1, 8 and 16 H100s on Marlowe, and measure time per step and GPU-hours at each scale.

What is here

Path	What
docs/	the website (Jekyll, just-the-docs), deployed by .github/workflows/pages.yml on every push to main
env/	paths and cache settings for the Yens (yens.sh) and Marlowe (marlowe.sh)
scripts/	setup, data preparation, labeling, scoring, step timing, and the Weights & Biases sidecar
slurm/train_args.sh	the training job every checkpoint runs
slurm/yens/	Yens training, data preparation, and serving a model to label filings
slurm/marlowe/	Marlowe training at 1, 8 and 16 GPUs

No data, results or model weights are in this repository. Jev's labeled filings are on the Yens at /scratch/shared/marlowe_hackathon/data/jev and on Marlowe at /scratch/m000106/gsb_hackathon/ceo-turnover/data/jev. .gitignore is a whitelist: only the paths above are tracked, and data files are ignored everywhere in the tree.

The models are pre-downloaded into a shared cache on each cluster. Models on the website lists them and how to download them again.

Build the site locally

cd docs
bundle install
bundle exec jekyll serve

Then open http://localhost:4000/marlowe-hackathon/.