nz00shuuuu/capture24-ts-haystack-fixed-needle

Capture24 TS-Haystack — Fixed Needle Length

Long-context retrieval / reasoning benchmark over Capture24 wrist-worn accelerometer recordings, used in Recursive Agents are Effective Time Series Reasoners (ARTS-RLM).

This repository supersedes nz00shuuuu/capture24-ts-haystack-cot for the paper's main capture24 experiments. Differences:

1. Fixed (absolute-ms) needle length of 3–10 s across every context length instead of needles that scale with context. With a 7200 s haystack the needle is now ~0.04–0.14 % of the signal (vs ~5 % under the ratio-based recipe), making longer contexts genuinely harder.
2. Signal sidecars — every data.parquet ships with an aligned signals.npy memmap (float16, 3-axis × samples), so the parquet itself stays small and the model loader can mmap signals directly.
3. Context lengths aligned with the LTAF / Sleep-PSG / UK-DALE haystacks (100 s, 900 s, 3600 s, 7200 s) for direct cross-dataset comparison.

Layout

.
├── tasks_absolute_needle/
│   ├── metadata.json                    # global generation config
│   └── {100s,900s,3600s,7200s}/
│       └── {task}/
│           ├── metadata.json            # per-task generation stats
│           └── {train,val,test}/
│               ├── data.parquet         # question/answer/info rows
│               └── signals.npy          # float16 sidecar (rows × samples × 3)
├── timelines/P001..P151.parquet         # per-participant activity timelines
├── bout_index.parquet                   # cross-participant bout index
├── transition_matrix.json               # activity transition probabilities
└── README.md

Tasks (10): existence, localization, counting, ordering, state_query, antecedent, comparison, multi_hop, anomaly_detection, anomaly_localization.

Each split has 1000 train / 150 val / 150 test samples per (context, task).

Download

Use the ARTS-RLM helper (recommended — mirrors the on-disk layout the configs expect):

python scripts/data/download_from_hf.py --dataset ts-haystack-fixed-needle

Or directly via huggingface-cli:

huggingface-cli download nz00shuuuu/capture24-ts-haystack-fixed-needle \
    --repo-type dataset \
    --local-dir data/capture24/ts_haystack

Regeneration

Re-deriving this benchmark from the raw Capture24 signals requires both this repo and nz00shuuuu/capture24-raw:

python scripts/data/download_from_hf.py --dataset capture24-raw
python scripts/data/download_from_hf.py --dataset ts-haystack-fixed-needle
python -m scripts.data.build_ts_haystack \
    --config src/datasets/ts_haystack/generation/absolute_needle.yaml

The timelines/, bout_index.parquet, transition_matrix.json shipped here are the exact artifacts used to draw needles and distractors — keep them in place if you want bit-identical regeneration.