UK-DALE-Haystack

A controlled additive-needle benchmark for long-context time-series language models built on top of UK-DALE (Kelly & Knottenbelt, 2015), the canonical UK domestic appliance-level + whole-house power demand dataset.

Each sample is a 6-second-sampled mains active-power trace with one or more real per-appliance bouts inserted at known locations. A QA prompt asks the model to detect, count, localize, order, or reason about those bouts across five context lengths from 15 minutes to 24 hours.

Quick stats

• 10 task families (existence, localization, counting, ordering, antecedent, comparison, multi-hop, state_query, anomaly_detection, anomaly_localization).
• 5 context lengths: 900 s, 3600 s, 7200 s, 32400 s, 86400 s (15 min, 1 h, 2 h, 9 h, 24 h).
• Per (task, ctx, split): 1000 train / 150 val / 150 test (≈65 K samples).
• 3 houses kept from UK-DALE (h1, h2, h5); 290 ISO-weeks split 80/10/10 by (house, week) at seed=42.
• Bouts and backgrounds are real; insertion is sample-by-sample additive (no style transfer, no blending, no scaling).

Layout

uk_dale_haystack/
├── manifest.json           dataset metadata (houses, appliances, regimes)
├── source_manifest.json    sha256 pin of the source UK-DALE HDF5
├── split_manifest.json     (house, ISO-week) → train/val/test
├── bout_index.parquet      one row per ON-event, split-pinned
├── bout_index_stats.json   per-appliance bout summary
├── tasks/{ctx}s/{task}/{split}/data.parquet   QA samples (metadata only)
├── signals/h{B}/m{M}.npy   float32 watts memmap per (house, meter)
├── signals/h{B}/m{M}.t.npy int64 unix-ns timestamps memmap (paired)
├── signals/conversion_manifest.json  HDF5→npy conversion provenance
└── diagnostics/            insertion-shortcut probe results (optional)

QA shards are lightweight metadata only — the mains trace for each sample is reconstructed at load time from background_house_id, background_start_ns/end_ns, and needles_json, by reading the npy sidecars in signals/. Both signals/ and tasks/ are required for use.

How to use

The accompanying codebase lives at the project repository (see src/datasets/uk_dale_haystack/) and provides loaders, samplers, and configs that consume this layout directly.

Download:

python scripts/data/download_from_hf.py --dataset uk-dale-haystack

This places everything under data/uk_dale/uk_dale_haystack/, ready for the runtime loader (src.datasets.uk_dale_haystack.qa_dataset.UKDaleHaystackQADataset). Training configs: configs/paper/{flamingo,chattime,chatts,itformer}_uk_dale_haystack.yaml.

Reproducibility

Every sample is generated with rng = default_rng(seed) where seed = md5(master_seed | task | ctx_s | split | sample_idx). Two generation runs with the same master_seed produce identical parquet shards (verified by tests/integration/test_uk_dale_reproducibility.py).

To regenerate from the source HDF5:

.venv/bin/python -m scripts.data.build_uk_dale_manifest
.venv/bin/python scripts/data/convert_uk_dale_to_npy.py
.venv/bin/python scripts/data/build_uk_dale_bout_index.py --overwrite
.venv/bin/python scripts/data/build_uk_dale_haystack.py --overwrite

The full design document, validation results (insertion-shortcut probe, classifier transfer, sim-to-real gap analysis), and limitations are in src/datasets/uk_dale_haystack/DATASET_CARD.md in the codebase.

Source & license

• Source dataset: UK-DALE — Kelly J., Knottenbelt W., "The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes," Sci. Data 2:150007, 2015.
• License: CC-BY-4.0 (inherited from UK-DALE; this derivative preserves the attribution requirement).
• Source pin: sha256 5cb9c07ec0da0ac4cc9e1a6ee6a7ba76561f4fe4367e6de78addaa2a63432092 on the NILMTK HDF5 export (5.896 GB).

Citation

If you use UK-DALE-Haystack, please cite the original UK-DALE dataset:

@article{kelly2015ukdale,
  title   = {The UK-DALE dataset, domestic appliance-level electricity
             demand and whole-house demand from five UK homes},
  author  = {Kelly, Jack and Knottenbelt, William},
  journal = {Scientific Data},
  volume  = {2},
  pages   = {150007},
  year    = {2015}
}