Private datasets
Two kinds of data are kept off the public download, and both are shown here as aggregate figures only — no case content. The held-out test set is synthetic data the platform team holds back so models can be evaluated on cases they cannot have seen. The study dataset is collected under ethics protocol BB 211/25 and held on the platform only as encrypted, never-decryptable blobs.
Held-out synthetic test set
Synthetic history-only vignettes for ICHD-3 chapters 1–4 (primary headaches), written by Claude models from reasoner-verified ICHD-3 graph skeletons and never published. Split by skeleton family: no public case shares a skeleton with a held-out case. Automatically verified, not clinically validated.
| Diagnosis group | English | German | Total |
|---|---|---|---|
| 1 Migraine | 66 | 65 | 131 |
| 2 Tension-type headache (TTH) | 36 | 34 | 70 |
| 3 Trigeminal autonomic cephalalgias (TACs) | 48 | 46 | 94 |
| 4 Other primary headache disorders | 36 | 33 | 69 |
| Total | 186 | 178 | 364 |
The hash is taken over the canonical file of the held-out cases, so the set can be verified when it is eventually released without being revealed before. Results on this set are produced by the platform team off-platform.
Study dataset (BB 211/25)
Study cases are held on the platform only as encrypted, never-decryptable blobs — the platform cannot read or count them, so these figures are maintained by the study team. The data itself is not publicly downloadable.
No study-dataset figures have been published yet.