News & updates
Dataset versions, platform launches, and schedule milestones — newest first. Subscribe to the low-volume mailing list to get them in your inbox.
- Final phase
Final-phase test inputs arrive with the Final on 19 October, and the dataset is final
Two things ahead of the Final (19–26 October). (1) The Final-phase test inputs will be published when the Final opens, at 00:00 UTC on 19 October, as a separate Hugging Face dataset, Sophelio/fusion-equilibrium-challenge-final, with two configs: diii_d_private_test (879 shots) and mast_private_test (1,205 shots). Like the public test split, they contain inputs only. The dataset you use now is not changed in any way and gets no new revision, so your local copies, caches and pinned revisions all stay valid. (2) There will be no v1.1.1 dataset. Dataset v1.1.0 together with data_fixes.py from the starter kit is the final data for this competition. The Final-phase inputs are built the same way, so data_fixes.py applies to them unchanged. This replaces our August note that a corrected dataset would follow. A full guide to the Final phase follows next week: what to submit, the reproduction package, and how results are reported.
- Platform
The Codabench Forum tab is hidden — use the direct link
Since Codabench's platform update on 29 September, the Forum tab no longer appears for participant accounts. The forum itself still works at https://www.codabench.org/forums/17429/, and questions, answers and organizer announcements are posted there. The same update makes the registration box say the competition requires approval; it does not, and registration is approved automatically. Codabench has fixed both, and the fix arrives with their next release.
- Clarification
Development feedback, which Final submission counts, honorable mentions, and CPU data generation
Four questions several teams asked, now answered in the rules. (1) The scores and Detailed Results the Development phase returns for your submissions may be used to compare, tune and select MAST pipelines; state that use in your methods report. Rule 3 bars MAST equilibrium data you can score against yourself, offline — not the platform's feedback. (2) Your most recent Final submission that scored counts for both awards by default; you may instead nominate one for Award #1 and another for Award #2 by email before the Final closes. For Award #2, G_ratio divides by your team's best DIII-D score across its scored Final submissions, so a weaker DIII-D model cannot raise it. A submission that fails to score never displaces an earlier one, and a slot lost to a platform fault is returned. (3) The four honorable mentions are judged on the MAST blind fold, using the submission that counts for Award #2. (4) Only GPU work counts toward the 24 GPU-hour rebuild ceiling; CPU-only data generation is declared but not capped, and generated data may ship with SHA256 checksums alongside the code that makes it.
- Rules update
MAST equilibrium data excluded from Challenge 2; reproduction packages required in the Final phase
Rule 2 permitted external public tokamak archives with disclosure. That was written with other machines in mind, and it collided with the premise of Challenge 2: the withheld MAST targets are drawn from the public FAIR-MAST archive, so "an external archive" and "the answer key" were the same object. Effective today, MAST/MAST-U equilibrium reconstructions from any source — FAIR-MAST, mirrors, derivatives, or model weights trained on them — may not train, fine-tune, calibrate, validate or select any component that produces MAST predictions, including hyperparameter and checkpoint selection, and including aggregates such as bases or priors. MAST inputs remain fully usable with disclosure: self-supervised pretraining, target-domain normalization, distribution alignment and label-free test-time adaptation are all permitted, as are MAST geometry, other machines' equilibria, and synthetic equilibria from your own solver. This also removes a question several teams asked — since MAST labels are out of scope entirely, there is no longer any need to certify that external MAST training data is disjoint from our test folds. Development entries are not judged retroactively; awards are decided on the Final phase. Separately: every Final-phase submission must carry a reproduction package reference at submission time (source, pinned environment, weights, a training entrypoint that rebuilds from scratch, an inference entrypoint, and per-machine data provenance), on any public forge or host — we rebuild the top five per challenge. Ties on the composite are now broken by R²ψ, then D_LCFS. Per-shot rows in Detailed Results are keyed by submission position rather than machine shot number, and per-shot detail is withheld entirely in the blind Final phase.
- Rules update
Challenge 2 eligibility gate raised: DIII-D composite S_model ≥ 0.85
A participant pointed out (starter-kit issue #7) that the Challenge 2 ratio G_ratio = S_MAST / S_DIII-D could be inflated by deliberately weakening a DIII-D entry: the old R²ψ > 0.6 gate constrained only the flux term — 55% of the composite — so driving DIII-D down to the eligibility floor bought roughly a 3× ratio boost without breaking any rule. Effective today, only entries with a DIII-D composite S_model ≥ 0.85 are ranked on Challenge 2. The ratio itself stays, because Challenge 2 is about zero-shot transfer: do genuinely well on DIII-D, then retain that performance on MAST. The gate is now enforced in the scorer — an entry below it shows G_ratio = 0 in the Ch2 column, with the reason stated in its Detailed Results. Challenge 1 is unaffected. Entries already at or above the gate keep their scores exactly and need not resubmit; below-gate entries keep their displayed ratio until their next submission, but Award #2 will be decided under the new rule either way.
- Data erratum
DIII-D plasma-current timestamps — fix shipped in the starter kit
A participant found that magnetics_plasma_current_times carries the wrong time axis on ~69% of DIII-D shots, so resampling Ip onto efit_times there reads pre-shot noise instead of the real current. The targets, scoring, and fairness are unaffected — this is one input column, and everyone has the same data. A verified drop-in fix (data_fixes.py) now ships in the starter kit and the baseline applies it automatically; v1.1.0 plus this fix is the final data for the competition (see the Oct 9 announcement), so nothing you build on it is throwaway. A small MAST Thomson channel-alignment fix ships alongside it.
- Now live
Training & public test data are live
The full 9,121-shot corpus (98 GB) is now on Hugging Face, opening Phase 1 (development). Every timeslice is a diverted plasma — limited frames were removed on both machines, so the boundary is always set by a magnetic X-point. Grab it from Resources, run a reference baseline, and make your first submission.
- Site
Challenge website is live
Browse the challenge motivation, dataset documentation, four reference baselines, the scoring breakdown, and the full timeline. Six sample shots ship in the starter kit for early exploration.