safe-diabetes-benchmark/transition_datasets

Safe Diabetes predictor transition dataset — rerun 2026-10-08

This release contains 15,120 accepted simulator collection tasks, 182,880 numeric NPZ files, and their unchanged JSON companions in 22 tar shards. The total selected native payload is 8,297,231,241bytes before tar headers. MIT licenses the released dataset and its included helper. This is synthetic research data for dynamics prediction; it is not patient records or a clinical dosing system.

Scope and source

Eight algorithms contribute policies: CPO, PPO-Lag, TRPOLag, RCPO, CRPO (stored as OnCRPO), PCPO, CUP, and FOCOPS. There are three diabetes configurations (T1D, T2D, T2D without pump), three age cohorts (child, adolescent, adult), three policy seeds (100,101,102), and separate deterministic/stochastic collection modes. CPO/PPO-Lag source provenance comes from historical training-log headers and retained file hashes, with no sealed full training-time source snapshot; the five other original algorithms and repaired PCPO use distinct frozen sources. Do not assume every policy was trained by the newest public code.

The policies were trained online in GlucoSim. This dataset was collected afterward to fit and validate the predictive-shield dynamics models. It is not the dataset that trained the RL policies. GlucoSim commit: ddd674e35a32a3a6406a0c2d5e7bf3c72f56dfec. See source-provenance.json and each portable task index for exact source, checkpoint/config, collection/archive hashes and seed lineage.

Splits

Every predictor split uses the corresponding age cohort's virtual patient #001 over a one-day prefix/episode. For each policy and collection mode, factual splits contain 24 training, 8 validation and 8 test episodes; branch splits contain 16 training, 8 validation and 8 test prefixes. Train/validation/test reset families remain separate. One branch prefix requests 11 anchors and 25 joint categorical candidates per retained anchor. Skipped anchors from an early-ended prefix are preserved with their original reasons; terminal trajectories are valid samples. No samples or skipped anchors were regenerated for publication.

The seven-day unseen-patient (#002–#010) outcome comparisons are a separate evaluation. Those trajectories are not included in this predictor dataset and were not used for predictor fitting or penalty selection. No unfinished native reward/cost supplement is represented by this release.

Numeric schema and units

schema.json records the exact14 observation features and their units/normalization. CGM and future CGM deltas use mg/dL, bolus IOB uses U, COB uses g, and CGM trend is the raw difference between successive five-minute interval-average CGM samples (mg/dL), not a per-minute rate. The cgm feature is the final-minute CGM reading, so differencing that column does not reconstruct cgm_trend. Other observation components are dimensionless time/meal/count features. Meal and bolus counts track accepted controller recommendations; autonomous meals are separate. Each factual NPZ contains the reset/current observation plus every subsequent observation, bolus/meal requested recommendation indices, accepted indices, acceptance flags, rewards, costs and termination/truncation flags. Branch NPZ files retain ragged candidate futures with explicit offsets, actual glucose deltas and policy-proposal metadata. All numeric files are loadable with allow_pickle=False; no pickled simulator state is included.

Recommendation indices 0–4 and simulator-accepted indices are not delivered physical insulin or carbohydrate doses. Rejected components become0. Basal insulin, autonomous scheduled meals and amount noise are separate simulator processes. Branch prefixes use no controller action; after the first forced candidate recommendation, future requests follow the declared fixed policy. Rewards/costs are the frozen simulator definitions and do not assert a clinical guarantee.

Loading

load_dataset.py locates a selected policy/mode collection, downloads only its shard from this repository, checks the shard/member hashes and extracts its unchanged native directory. See its --help for local and Hub usage. Use the GlucoAlg causal-episode/branch loaders for scientific semantics; the corrected public loading source is commit 2eaed2af25a0035365d9112c7b4083a31934b802 of GlucoAlg. The loader source hashes used to validate publication are recorded in source-provenance.json.

Filter by algorithm, patient type, cohort, policy seed and action mode before passing native collections to a loader. Reset and anchor IDs recur across different policies/modes; use the full portable source/policy/mode/task/member identity as the dataset key. Original runtime path strings in unchanged JSON are historical metadata, not fetchable download paths.

Integrity and limitations

CPU jobs authenticated the retained stage inventories and every task proof/archive, checked every archive/member hash, loaded every NPZ with pickle disabled, validated native schemas and split identity, and checked written shard round trips. SHA256SUMS, shards.jsonl, index/*.members.jsonl and index/*.tasks.jsonl expose portable integrity and lineage. Internal H1 bookkeeping is not an independent public attestation. Structural integrity does not establish physiological realism, predictive calibration or shielding benefit.

The prior 180 NPZ Hub files are a separate legacy release and must not be relabeled as this rerun. The associated paper's May 2026 version predates this dataset; these September–October 2026 rerun artifacts have their own source/protocol manifests.