Track 01 · Tech≤ 50M params
Foundational LLM
Build and train a language model capped at 50 million parameters, from scratch — scored on HellaSwag, ARC-Easy, PIQA, WinoGrande and WikiText-103 perplexity.
Explore the track
Challenge tracks
Three tracks. Pick one.
Each team submits exactly one project to exactly one track, and is scored only against that track’s criteria. Judges may reassign or disqualify a project submitted to the wrong track.
Teams of 1–6One project, one trackWorking code required
Track 02 · AppliedReal, de-identified data
Med Tech & Finance
Predictive models, diagnostic pipelines or automated financial systems on real, empirical data — research prototypes, never deployed on real patients or real money.
Explore the trackTrack 03 · OpenNo limits
General invention
A working prototype with technical novelty and cross-disciplinary engineering — software, hardware, interactive media or generative systems.
Explore the track(01) Side by side
Compare the tracks.
| Track | Build | Hard rule | Judged on |
|---|---|---|---|
| 01 · LLM | A language model trained from scratch | ≤ 50M trainable parameters incl. embeddings & output head | Perplexity & accuracy · reasoning · training efficiency · innovation · documentation & demo |
| 02 · Med/Finance | Predictive ML, diagnostic pipeline or automated financial system | Public or de-identified data; no live deployment | Innovation & impact · technical feasibility · rigor & validation · presentation |
| 03 · Open | A working prototype with technical novelty | No architectural or resource limits | Creativity · execution · impact · presentation |
All tracks: no theoretical papers, design documents or slide decks alone — there must be running code, a working system or a functioning physical build.