WMT·Hackathon rules
Rules
Hackathon “Can a small model pass the Polish history matura?” · Kolektyw3, Warsaw · 25–27 September 2026
The essentials — read this if nothing else
All submissions go through our app: warsawmodeltrainers.dev/matura. Step by step:
- Create an account with the same email you used to register on Luma — otherwise we can’t link you to your registration.
- Create a team or join an existing one (email invite). All members must be on the team list — register your team lineup in the app by Saturday 16:00.
- Register your base model: name, link to the weights, number of parameters (max 12B).
- Submit the untouched model to the leaderboard — this is your benchmark. For your result you submit both models, the untouched one and the trained one, so we can compare progress. Without a benchmark you can’t compete on progress.
- Submit new versions whenever you like (max once per hour). The screen in the room shows how every team is improving.
- Sunday 11:00 — stop. You freeze your model and harness; the app no longer accepts submissions. Report technical problems by 11:15 — later reports won’t be considered.
Exam and presentation on Sunday, from 11:00 on stage. The app draws the order, shown on screen from 10:45. Your model’s score on the final exam sheet appears on screen, then you have 2–3 minutes to talk about the problems you hit while training and 1–2 minutes for jury questions. Polish or English. Bring your own laptop with HDMI, or a model ready to run on our station. You can leave the queue and come back before your turn.
Questions about the app and anything else: our Telegram group with organizers, sponsors and mentors — t.me/warsawmodeltrainers.
General
Is the hackathon free? Yes. Participation, sponsor compute, food and the closing party are free.
How do I sign up? Through Luma. We review applications and confirm by email. There are 60 places — your Luma confirmation is your ticket.
Key times:
| Fri 25.09, 15:00 | Harness-building workshop (optional) + registration |
|---|---|
| Fri 18:00 | Opening: rules, sponsors, compute, how to pick a model (data sources: all are fine) |
| Fri 20:25 | Team building — find a team here if you don't have one |
| Fri 21:05 | Work starts, leaderboard opens |
| Sat, by 16:00 | Register your team lineup in the app — the same one where you submit your project |
| Sat 12:00–18:00 | Mentor slots (book in the app) |
| Sun 11:00 | Training ends, exams and presentations start |
| Sun 13:00 | Panel “Builders vs. Enablers” |
| Sun 14:00 | Results and prizes, then the Smok x Badideas closing party until 22:00 |
Is there an official chat? Yes — our Telegram group: t.me/warsawmodeltrainers. Organizers, sponsors (including compute: Nebius, Labqoat) and mentors are there, plus announcements and help with the app.
Do I need to know how to train models? No. You can skip training entirely — wrap a model in a harness (RAG, prompts, tools) and still sit the exam. The Friday 15:00 workshop is about exactly that.
How do I get a mentor? Mentors aren’t assigned in advance. On Saturday 12:00–18:00 there are two mentor tables (harness & evaluation / training & data) — book a 20-minute slot in the app. JetBrains mentors walk around the room. Nebius and Labqoat help with compute: Friday 21:00–23:00 at the table by the entrance, then on Telegram. We strongly recommend one mentor slot on Saturday — teams that take one usually finish with a better score.
Who can take part, and teams
Requirements: you are at least 18 on the last day of the hackathon; you are registered and confirmed on Luma; you are on site — this is an in-person hackathon, not hybrid. Every team member registers separately and picks up a wristband at the entrance.
Do I need a team? No. You can compete solo or in a team of up to 4. If you don’t have a team, come to team building on Friday at 20:25: anyone with an idea gets 30 seconds at the mic, then a sticky-note board (I have an idea / looking for a team / I can do X) and 15 minutes of mixing.
Does everyone have to be there all weekend? No — but at least one team member must be at the exam on Sunday, and your team lineup must be registered in the app by Saturday 16:00. We close the space at midnight — overnight training runs remotely on sponsor compute.
Can I be in two teams? No. One person, one team, one submission.
The task and the rules
Task: take any open model up to 12B parameters and make it score as high as possible on the Polish history matura. You may train (fine-tuning, LoRA, further training on your own data), wrap it (RAG, prompts, agent, tools), or combine both. What counts is the exam score and how much the trained model improved over the untouched one (your benchmark) — which is why you submit both.
| Base model | any model with downloadable open weights, up to 12B parameters (e.g. Bielik, PLLuM, Qwen, Gemma, Llama). Register it in the app: name, link, number of parameters |
|---|---|
| Size limit | up to 12B parameters. A RAG knowledge base doesn't count towards the limit; the repo contains either the base itself (if the source licences allow it) or a script that builds it |
| Benchmark | the score of the untouched base model. For your result you submit both models — untouched and trained — and progress is the difference. The app checks the checksum of the untouched weights; a trained model doesn't count as a benchmark. You may switch models during the event: on Sunday the benchmark of the model that sits the exam counts |
| Leaderboard | from Friday 21:05 to Sunday 11:00, max one submission per hour per team. Questions are drawn from a large pool for each submission. Leaderboard scores don't count towards the result — only the benchmark and the final exam do |
| Training data | you find it yourself — any source is fine as long as you have legal access to it (including synthetic data from any LLM; closed APIs are fine here). See “Data sources and copyright”. Datasets derived from Wikipedia are published under a CC BY-SA-compatible licence |
| Exam | the model answers on its own. During the exam the harness may not call a closed API (GPT, Claude, Gemini etc.) or a second model beyond the limit. No internet during the exam — RAG on a local knowledge base only |
| Final exam sheet | separate from the practice sheet; teams see it only on Sunday at 11:00. Any attempt to get it earlier = disqualification |
| Compute | credits from Nebius (a credit code for every participant) and Labqoat ($2,000 in total, split equally between teams); instructions in the starter repo. Anyone can also open a free Modal account ($30 of compute per month on the Starter plan, modal.com/pricing) — Modal is not a sponsor. Your own GPUs are allowed |
| Own work | everything we judge is created during the hackathon (from Friday 18:00). You may use public libraries and your own pre-existing tools, but a model trained before the hackathon can't compete |
What your final submission must include
Your submission is complete when, before Sunday 11:00, the app contains:
- the full list of team members (lineup registered by Saturday 16:00);
- both models: the untouched base model (benchmark) and the frozen trained model (weights, or weights + adapters) with the harness that sits the exam — exactly what you will run on stage;
- a link to a GitHub repository (public, or private with read access for the organizers and the jury) with the training and harness code, data scripts, list of sources and a README that lets someone reproduce the result;
- a SOURCE.md file in the repo with the sentence: “Made during the Warsaw Model Trainers hackathon, Kolektyw3, 25–27.09.2026”. Without this file a team can’t win any prize;
- code under a licence of your choice; weights under the base model’s licence; Wikipedia-derived datasets under CC BY-SA or a compatible licence.
You can edit your submission until 11:00 on Sunday. What isn’t in the repo, the jury doesn’t see — talking about it on stage doesn’t replace the code.
Ownership and licensing
Everything you build during the hackathon — code, harness, prompts, tools — belongs to your team, and you can use it afterwards however you like. Your repo can be public or private; if it’s private, give the organizers and the jury read access before Sunday 11:00 (we only use it to judge your submission). You choose your own licence.
With your team’s consent, we may publish the winners’ models and datasets on Hugging Face after the event, where the source licences allow it.
Data sources and copyright
All data sources are allowed, as long as you have legal access to them. In return, three rules apply:
- Models and data built from third-party content are non-commercial. Models and datasets built from third-party content may only be used for educational or research purposes, unless the source licences allow more. This does not apply to your own code and harness.
- No circumventing protections. You don’t break paywalls or logins, and you don’t download content from sites whose terms forbid it.
- No republishing other people’s content. You don’t put copyrighted material in full into your repo (e.g. textbook scans or other people’s datasets without a licence). Instead, you provide a list of sources with links and a script that downloads and processes the data. Openly licensed content (e.g. Wikipedia — CC BY-SA) is published in line with its licence.
The team is responsible for its choice of sources. The organizer may ask for material to be removed from the repo, or exclude a project that breaks these rules.
Scoring, jury and prizes
How is the result calculated? Three parts. Score on the final exam sheet — 40%, calculated automatically. Progress of the trained model over the untouched one (benchmark) — 40%, also automatic. Presentation — 20%, scored by the jury right after you present, on a card with three questions: do you know what worked and why; quality of data and method; can it be reproduced. A small model with a big jump can beat a big model that stood still.
Who is on the jury? Gleb Badzeika (JetBrains), Jacek Migdał (Quesma) and Paulina Jóśków (Kolektyw3). The jury sits in the front row through all the exams, deliberates 13:00–13:45 and decides only on presentation scores and special prizes. The jury’s decisions are final.
Prizes. Main prize: PLN 8,000 for 1st place (paid by Kolektyw3) + $200 in Cursor credits per member of the winning team, to use after the hackathon. Every participant also gets $200 in Cursor credits for the hackathon — pick up your code from the organizers on Friday. Quesma special prizes (outside the ranking): Tokenmaxxer — most tokens used, measured in API-equivalent $; Token diversity — most different models used for vibecoding. Quesma will present the details on stage on Friday. Sponsors hand out the prizes on Sunday at 14:00; cash prizes and credits are paid out by whoever funded them.
Disqualification: a closed API in the harness during the exam, a model over the limit, no untouched-model (benchmark) submission for the model that sits the exam, commercial use of models or data built from third-party content, breaking the data-source rules, trying to obtain the final exam sheet, a model trained before the hackathon, no SOURCE.md. The organizer settles disputed cases on site.
Venue, hours, food, code of conduct
Where? Kolektyw3, Koszykowa 54, 3rd floor, left from the lift. Warsaw.
Opening hours: Friday 15:00–24:00, Saturday 9:00–24:00, Sunday 9:00–22:00. We close overnight — no sleeping in the space. Run overnight training remotely on sponsor compute and come back to your results in the morning.
Food: breakfast on Saturday and Sunday from 9:00, lunch on Saturday at 14:30 and Sunday at 12:00, pizza on Friday and Saturday evening, coffee and water all the time. Allergies and dietary needs — tell us on Telegram by Friday morning.
Equipment: bring your own laptop and charger. WiFi, extension cords, a 65″ screen, microphones and the exam station are on site.
Code of conduct: the ETHWarsaw Code of Conduct and the Kolektyw3 Space Rules apply. In short: respect people and equipment; report problems to the organizers on site or at paulina@ethwarsaw.dev. A breach means being asked to leave the space, no discussion.
Photos and materials: we take photos and videos during the hackathon; by registering you agree to their publication on the channels of Kolektyw3, AI Tinkerers and the sponsors. For publishing the winners’ models and data, see “Ownership and licensing”.
Contact details for sponsors: after the hackathon we will share your first name, last name and the email address from your registration with the event sponsors, so they can contact you about hiring, programmes and offers for participants. By registering, you agree to this. If you don’t want your details shared, email paulina@ethwarsaw.dev before the end of the hackathon and we won’t pass them on.
Back to the site