neuralsorcerer/thermoshift-1b

ThermoShift

Hourly synthetic building-cooling trajectories with observed decisions and paired counterfactual outcomes.

The dataset contains three cooling actions, their logging probabilities, factual outcomes, and an oracle table with latent state and potential outcomes for every action. Buildings are assigned to train, validation, and test splits, including heatwave and sensor-degradation conditions.

Configurations

Name	Contents
logged	Decision-time observations, cooling actions, propensities, factual targets, and trajectory boundaries
oracle	Latent temperature, physical parameters, sensor labels, three potential outcomes, and the best one-step action

Join configurations using row_id; building_id and step identify the position within each trajectory. schema.json describes every column, type, unit, and role. feature_roles.json supplies the policy and transition-model feature lists.

Loading

Use the repository ID shown on this dataset's page:

from datasets import load_dataset

records = load_dataset(
    "neuralsorcerer/thermoshift-1b",
    name="logged",
    split="train",
    streaming=True,
    columns=["row_id", "building_id", "step", "obs_temp_last_c", "action", "y_next_temp_c"],
)
print(next(iter(records)))

Set revision to a completed commit SHA when recording an experiment. Streaming reads records as they are consumed. Keep trajectories ordered for sequential tasks and match oracle labels by their identifiers.

Tasks

• Predict next-hour temperature, energy, emissions, or comfort deviation.
• Estimate action effects using the paired outcomes as evaluation labels.
• Evaluate one-step policies on logged states with the supplied propensities.
• Reconstruct sensor state and identify drift or missing observations.
• Measure generalization across buildings and the two stress conditions.

Policy inputs come from policy_features. Transition predictors can also use the logged action. Oracle fields provide evaluation labels or an explicitly selected supervision source. One-step potential outcomes branch from the current factual state; evaluating a different trajectory policy requires simulating its resulting state sequence.

Model and provenance

A single-zone thermal model advances each building by one hour. Cooling levels are 0%, 50%, and 100% of rated electrical input. Weather, occupancy, grid availability, and sensing evolve over the trajectory. Electricity and comfort determine the reward. DATASHEET.md gives the equations and parameter distributions.

run_config.json records the configuration, generator fingerprint, and runtime versions. manifest.json contains counts, sizes, and file checksums. The completion marker binds the manifest and release metadata. validation.json records the most recent saved validation status and its check scope.

License

The dataset is distributed under the MIT license included in LICENSE.

This release

1,000,000,000 unique hourly decision records from 5,952,381 synthetic buildings. Each decision has a matching row in the logged and oracle configurations, linked by row_id.

Split	Decision records
test	100,018,296
test_heatwave	50,108,016
test_sensor	50,205,960
train	699,746,200
validation	99,921,528

Seed: 42. Episode length: 168 hourly steps. Hidden-confounding mode: False. Config fingerprint: b32ebac0c5e6da38f1a7aadad9c77ff4388e8f67d8c05186d4f0b09a3c9d4538.

Compressed Parquet bytes (both configurations): 199915309678. Generation parameters and runtime versions are in run_config.json. Column roles and units are in schema.json.

Reproduction and validation

Generator: ThermoShift commit 3d92172f8168. The generator and its physical model were used unchanged.

reproduction/publish_streaming.py generates complete building shards, runs the upstream record and equation checks on every row, and verifies uploaded file sizes and SHA-256 hashes before removing local data. Per-shard validation receipts are in provenance/shards/. Finalization checks exact coverage of all configured shard intervals and both tables. validation.json contains the aggregate results from generation-time validation. When present, audit/validation.json records the separate full download, equation scan, and exact deterministic replay audit; _AUDIT_SUCCESS.json identifies its successful completion.

Use the completed immutable revision recorded in your download or experiment. _SUCCESS.json binds the final manifest and dataset documentation.