Quome/openworld-diagnosis

OpenWorld · Diagnosis Worlds (World-Time Compute)

A parametric family of diagnosis POMDP world-models for world-time compute — learn the diagnostic skill on many verified worlds, then test on entirely unseen specialties (whole worlds held out, scikit-learn style).

Load it

from datasets import load_dataset

ds = load_dataset("Quome/openworld-diagnosis", "hard")  # or "easy" / "world_count"

---

A parametric family of diagnosis world-models for studying world-time compute — whether fine-tuning a model on trajectories traversed through many verified world-models makes it generalize to unseen worlds from fewer real examples (OpenWorld experiments E74–E76; paper §"World-time compute").

What it is

Each specialty is a diagnosis POMDP: a hidden disease d; binary features (symptoms/tests) with disease-specific emission probabilities M[d, f]; the task is to read the specialty's disease profiles and a patient's observed features and name the disease. All specialties share one template (and action grammar); they differ only in their symptom→disease structure M and disease prior. The diagnostic skill (match a patient to the profiles) is therefore transferable across specialties, while each specialty's facts are specialty-specific — which is exactly what makes "learn the skill on train specialties, test on held-out specialties" a meaningful generalization measure.

Provenance — generated, not authored

These specialties are deterministic, seeded numpy parameterizations — NOT produced by Claude Code or any LLM. make_specialty(seed) draws a random emission matrix M (disease × feature) and a Dirichlet prior from a fixed RandomState(seed), so every specialty is reproducible bit-for-bit from its integer seed. (This is distinct from the 100 sector recipes elsewhere in OpenWorld, which were authored by Claude Code from one-line prompts. This family is synthetic-parametric on purpose: controlled, infinitely scalable, and equipped with a clean ground-truth oracle.)

Splits (scikit-learn style: whole worlds are held out)

The test set is composed of entire specialties never seen in training — generalization means competence on new worlds, not new states of a familiar world.

Family	Diseases × Features (sig)	Sig prob / base rate	Bayes oracle	Prior floor	Train / Test specialties
easy/ (E74)	6 × 12 (3)	0.70–0.95 / 0.10	~0.86	~0.34	60 / 20
hard/ (E75)	10 × 20 (4)	0.58–0.78 / 0.18	~0.69	~0.23	60 / 20
world_count/ (E76)	hard params	—	~0.69	~0.23	8…512 / 20 (fixed hard test)

world_count/ varies only the number of train specialties (sft_train_N{8..512}.jsonl) against a fixed held-out hard test, to measure whether more worlds traversed keeps lifting held-out accuracy. Train seeds are disjoint from the test specialties in every split.

Schema (JSONL)

• sft_train*.jsonl: {"prompt": <profiles + patient features + instruction>, "completion": "disease_k"}
• test.jsonl: {"prompt": ..., "answer": "disease_k", "specialty": <int test-specialty id>}

The prompt is self-contained: it states the specialty's disease profiles and the patient's present features, and asks for a single disease_k.

How to use

Fine-tune on sft_train, evaluate diagnostic accuracy on test (held-out specialties). Reference points: the oracle (full-information Bayes with the true M) is the ceiling; prior-only (guess the most likely disease) is the floor. Both are computable offline.

Reproduce

python datasets/openworld-diagnosis/generate.py

Deterministic; re-runs the committed seeded generators (experiments/e74_data.py, e75_data.py, e76_data.py) and rematerializes every split identically. The full meta-learning / oracle analysis is experiments/e74_diagnosis.py.

Results & paper

Evaluation artifacts live in experiments/results/ (e74_scaling.json, e74_diagnosis.json, e75_artifacts/, e76_*, e74_gemini.json) and are written up in the paper's "World-time compute" section.

Caveats

Synthetic data — not real clinical records. The oracle is Bayes-optimal on the generated M, so accuracies are bounded by the (deliberately tuned) task difficulty, not by medicine.

License

Apache 2.0 (same as the OpenWorld repository).

---

From the OpenWorld project

This dataset is produced by OpenWorld — a framework for verified symbolic world models, where a world's dynamics are explicit, auditable Python code (no training, no GPU). "World-time compute" is the idea that traversing many verified worlds of a domain and fine-tuning on that experience makes a model generalize to unseen worlds from fewer real examples.

• 💻 Code, experiments & paper: https://github.com/quome-cloud/openworld

Citation

@software{openworld_diagnosis_2026,
  title  = {OpenWorld · Diagnosis Worlds (World-Time Compute)},
  author = {Schwoebel, Jim},
  year   = {2026},
  url    = {https://github.com/quome-cloud/openworld},
  note   = {Hugging Face dataset: Quome/openworld-diagnosis}
}