Marlowe Hackathon 2026

A one-day, hands-on hackathon for Stanford GSB researchers, run by DARC and the HAI Marlowe team. You pick one of three tracks: each starts on the Yens and then fine-tunes an open-weight model across GPUs and nodes on Marlowe, Stanford’s H100 cluster. You leave with a working Marlowe job, Slurm templates for your own research, and the measurements a Marlowe project application asks for.

Friday, November 13, 2026 · 10:00 am to 5:00 pm · GSB, in person


The project: CEO turnover in SEC filings

When a public company’s CEO leaves, the company files an 8-K under Item 5.02. A large hosted model, Jev, has read every Item 5.02 filing since 2004, about 300,000 of them, and answered three questions about each: what happens to the CEO, why, and whether the change was announced in advance.

You will work with open models from the Kev family that answer the same questions. Training a Kev model on Jev’s labels, including how confident Jev was, is distillation: a model you run yourself can then label filings at scale on a GPU you control.

Before training, Kev-4B agrees with Jev on 58% of CEO changes. During the day you train, label filings at scale, and measure what each costs on the Yens and on Marlowe. The task and the data has the details.


Tracks

Pick one track. Each task ends in a check-in: post your result in the Slack channel and put up a 🟩 green sticky, or a 🟥 red one if you are stuck.

Task Beginner Intermediate Advanced
1 Score Kev-4B before training (Yens, 1 GPU) Fine-tune and profile Kev-4B (Yens, 1 GPU) Score Kev-27B before training (Yens, 1 H200)
2 Fine-tune Kev-4B (Yens, 1 GPU) Move your data with Globus (Yens to Marlowe) Fine-tune Qwen 27B (Marlowe, 1 H100, then 8 H100s)
3 Label filings on one Marlowe GPU (Marlowe, 1 H100) Label the test split on a node (Marlowe, 1 node) Fine-tune on two nodes (Marlowe, 16 H100s)
4 Label on a full node (Marlowe, 1 node) Label the corpus on two nodes (Marlowe, 2 nodes) Inference at scale with Qwen 235B (Marlowe, 8 H100s)
5     Distill a student (Marlowe, 1 H100)

Schedule

Time Session
10:00 Welcome: goals, the three tracks, the Slack channel, a quick login check on the Yens and Marlowe
10:15 Talk: an introduction to Marlowe (Marlowe team)
10:50 Talk: a social science researcher’s Marlowe story
11:30 Break; open a terminal on the Yens
11:40 Hands-on 1: task 1 of your track
12:30 Lunch
1:15 Hands-on 2: task 2
2:00 Hands-on 3: task 3
2:30 Break
2:40 Hands-on 4: task 4
4:00 Awards, templates, next steps for applying

Take it home

Every script here is a template you can adapt: swap in your own documents, questions and labels. Take It Home shows what to change.


Page last modified: Oct 30, 2026.

Your details

Set your SUNet ID once and the commands on every page fill it in. Stored in this browser only.

Built with Just the Docs.