Ziquan/cohort-world

CohortWorld data

Group-level data derived from primary-care and hospital records, for the CohortWorld environment and benchmarks. Use of the data is subject to the data provider's access conditions and to the licence below. Code, environment and reproduction scripts: https://github.com/Chrisa142857/cohort-world.

Every released cohort holds at least 10 participants, its size is published only as a band, and no individual-level record is included.

Labels are hidden

The incidence labels are not public. Only 20 cohorts keep their rate tables: the cohorts listed in official/labelled_cohorts.json, all from the largest size band (1,000+). In those tables a cell is published only if the rate times the lower bound of the cell's at-risk band is at least 5 events, rounded to 2 significant digits; every other cell is null. Every other cohort has "label": "hidden" and null rates.

As examples, the first 20 rows of the CohortCal train split keep their answers ("label": "published"): questions on cohorts of 1,000+ whose rate times 1,000 is at least 5 events, rounded to 2 significant digits. Their original row positions are in official/labelled_cohorts.json. The answer of every other rate or direction question is hidden (null).

To evaluate a model or to enter the leaderboard, send the model's predictions to [email protected], and we will score them and return the results:

• CohortCal: one JSON line per row of cohortcal/PhenoCal_C2_test.jsonl, {"idx": <0-based row index>, "pred": <rate in [0, 1]>}.
• CohortThink: one JSON line per item of items/t3_rate_unseen.jsonl, {"item_id": "...", "pred": <rate in [0, 1]>}.

Please include the model name, the prompt or agent setting, and the decoding settings. Requests for labels for other research are handled the same way.

Contents

directory	dataset	size
official/	the cohort collection	29,884 cohorts: 20,180 CohortWorld cohorts and the 9,704 cohorts behind CohortCal
cohortcal/	CohortCal, the calibration benchmark (2-character outcome codes), answers hidden	32,120 / 6,474 / 31,784 questions (train / validation / test)
items/	the CohortWorld task items, answers hidden	see below
paper_outputs/	the single-turn model predictions behind Tables 3 and 4, without the true rates	12 + 2 files

official/, the cohort collection

CohortWorld cohorts at three coding depths. The depth sets the outcome codes a cohort's rates are reported for: 19 ICD-10 chapters (C1), 183 two-character codes (C2), 1,374 three-character codes (default).

depth	cohorts	train / val / test
C1	5,619	2,892 / 561 / 2,166
C2	7,113	3,615 / 695 / 2,803
default	7,448	3,645 / 745 / 3,058

file	contents
CohortWorld_{depth}_{split}_states.jsonl	one cohort per row: state_id, its axes (age band, sex, diagnosis / procedure / prescription history), the rendered description, n_band, label (published or hidden), and rates, the cumulative incidence of a first diagnosis at horizons 1, 2, 3, 5 and 10 years for each outcome code (also at each outcome depth in rates_g1, rates_g2, rates_g3), null when hidden
labelled_cohorts.json	the 20 cohorts that keep their rates, the publication rule for their cells, and the original row positions of the 20 CohortCal example rows
CohortWorld_{depth}_manifest.json	horizons, outcome codes, and the rate definition
CohortWorld_{depth}_diseases_g{1,2,3}.json	the outcome codes that index rates_g1, rates_g2, rates_g3
CohortWorld_{depth}_vocab.json	the codes an agent may add or switch to
CohortWorld_{depth}_transitions.jsonl	the edits that lead from each cohort to another
CohortWorld_legal_moves.jsonl	the moves each cohort accepts, with their cost and the destination's size band
CohortWorld_{depth}_atrisk_bands.jsonl (+ .idx)	for each cohort, the banded number at risk behind each (horizon, outcome) cell; .idx holds byte offsets for seeking
CohortWorld_outcome_split.json	the outcome chapters held out as unseen
PhenoCal_C1_axes_sidecar.jsonl, PhenoCal_C2_axes_sidecar.jsonl, PhenoCal_axes_sidecar.jsonl	the CohortCal cohorts at each depth (3,640 / 3,198 / 2,866): size band, plus the procedure (OPCS-4) and prescription (BNF) codes carried by at least 10% of members, each with its share
marginal_incidence_landmark_gplinked_3ver.json	the population incidence of each code and horizon at each depth, the b of the CohortThink AUROC; suppressed as described below

cohortcal/

PhenoCal_C2_{train,val,test}.jsonl. Each row is a written history with a disease and horizon (prompt), its answer (ground_truth.rate, null except in the first 20 train rows), and metadata (disease, horizon, the cohort's size band). Splits are profile-disjoint: no cohort description appears in more than one split.

items/

task	train	seen	unseen
CohortThink, rate (t3_rate_*)	16,770	880	943
locate (t1_locate_*)	16,589	870	919
delta (t2_delta_*)	297	60	108

Each item is a question in words (description, ask), the cohort it was drawn from (source), and its answer (answer). The rate of a CohortThink item and the direction of a delta item are hidden (null); the cohort a locate item asks for is given. The paper's Table 4 uses the 943 unseen CohortThink items.

paper_outputs/

table3/ holds one record per single-turn row of Table 3 (the ten open models and the two vendor APIs): rows of {idx, disease, horizon, pred, text}, the model text verbatim. table4/ holds the two no-tool rows of Table 4. The true rates are removed, and the agent traces are not released, because the tool results in them show cohort rates. Each directory's index.json maps the table row to its file, row count and SHA-256.

Checkpoints

The trained policies are separate model repositories: Ziquan/cohortworld-qwen3.5-9b-{sft,csgrpo,csgrpo-psi} and Ziquan/cohortworld-qwen3.6-27b-{sft,csgrpo,csgrpo-psi}, Apache-2.0, bf16 as exported.

Using the data with the code

hf download Ziquan/cohort-world --repo-type dataset --local-dir data
export CW_PANEL=$PWD/data/official CW_OFFICIAL=$PWD/data/official CW_SIDECAR=$PWD/data/official
export CW_PHENOCAL=$PWD/data/cohortcal CW_ITEMS=$PWD/data/items
export CW_MARGINAL=$PWD/data/official/marginal_incidence_landmark_gplinked_3ver.json

The environment reads a cohort's rate when an agent observes it, so with the labels hidden it can serve observations only for the 20 labelled cohorts. Scoring needs the hidden answers; send predictions as described above.

Disclosure control

• Every cohort has at least 10 participants, and its size is published only as a band.
• Cohort rates, question answers, event counts and exact cohort sizes are not published, except the rate tables of the 20 labelled cohorts and the answers of the 20 CohortCal example rows, which follow the rules above.
• In the marginal incidence table, a cell in which at least one but fewer than 10 distinct participants had an event is null. Every other cell reports only the rate, rounded to 2 significant digits. Event, at-risk and participant counts are not published.

Not released: the individual-level evaluation items (one participant each), any count behind a band or rate, the supervised cold-start demonstrations (their answers state rate differences), the train-split mean rates, and the agent traces.

Licence

The data are released under CC BY-NC 4.0 and remain subject to the data provider's access conditions stated at the top of this card. The code in the GitHub repository is Apache-2.0.

Ethics and scope

A rate here is the incidence of a group a person matches. It is not a personal prognosis, it carries the selection of the source cohort, and it needs recalibration before use on another population.