nz00shuuuu/anomalyxl-generalization

AnomalyXL Generalization

Real-data, out-of-distribution generalization sets for the AnomalyXL time-series anomaly task, from the paper TimeRLM: Recursive Language Models Are General Temporal Reasoners.

Each row is a single, isolated anomaly (or a clean negative) spliced from a real clinical recording and cast into the exact anomalyxl-precise classify_with_evidence schema, so the timeseries_qa environment scores it unchanged. Because no synthetic signal appears here, performance on these sets measures zero-shot transfer of a recipe trained on AnomalyXL to genuine physiological domains.

The gold span is small relative to its window — a median of 1.7 s of ECG and 17.7 s of sleep signal — so localization is a genuine needle-in-a-haystack task.

Subsets

ltaf-anomaly/ — LTAF ECG (128 Hz)

• Env name: ltaf_anomaly | channels: ECG_0, ECG_1
• Rows: 250 (200 anomaly / 50 no-anomaly)
• Anomaly: a single premature beat (PVC / APC) in normal sinus rhythm. Gold window: the RR bracket (previous → next beat) around the ectopic beat.
• Kind balance: No anomaly 50 · Premature Atrial Contraction (APC) 99 · Premature Ventricular Contraction (PVC) 101
• Window length: 60–1017 s (median 332 s), i.e. 7,691–130,228 samples.

sleep-anomaly/ — Sleep-PSG (200 Hz)

• Env name: sleep_anomaly | channels: AIRFLOW, CHEST, ABD, SaO2
• Rows: 250 (200 anomaly / 50 no-anomaly)
• Anomaly: a single respiratory arousal bout (Central Apnea / Obstructive Apnea / Hypopnea / RERA). Gold window: the bout's native [start, end) range.
• Kind balance: Central Apnea 50 · Hypopnea 50 · No anomaly 50 · Obstructive Apnea 50 · RERA 50
• Window length: 60–640 s (median 313 s), i.e. 12,028–128,003 samples.

Construction

Every positive window is drawn so that it contains exactly one anomaly: the window is sampled inside a corridor clipped at the nearest neighbouring event, which guarantees no other event intrudes. 80% of rows are positives (carrying the anomaly kind and a [start, end) gold window); 20% are negatives (kind No anomaly, no window). Channels are per-channel z-normalized, and a single unified detect-and-classify prompt is used for both positives and negatives.

Windows span two length regimes — short (60–300 s) and long (≥ 300 s, stratified so every anomaly kind is equally represented, capped at ~131k samples to fit a reasoning model's context budget). Draws are seeded (42 for the short windows, 43 for the long ones), so the set is reproducible from the raw PhysioNet recordings with the builders released alongside the benchmark (anomalyXL/generalization_set/).

Schema

Columns follow anomalyxl-precise. Key fields:

• answer — "classify_with_evidence:" + json({"kind": <label>, "start": <int>, "end": <int>, "_L": <int>}); for negatives the payload is {"kind": "No anomaly", "_L": <int>} (no window).
• channel_names / channel_values — the z-normalized series (list of channels).
• length, n_channels, question, injection_kinds/channels/starts/ends, seed.

Scoring: kind_x_iou = (kind matches) × IoU of the predicted vs gold [start, end); a correct No anomaly prediction is credited on negatives.

Usage

The timeseries_qa environment takes either a local directory or a Hub shard as repo_or_path, where a #fragment selects the subset:

vf-eval timeseries_qa -m <model> \
  -a '{"dataset":"ltaf_anomaly","repo_or_path":"nz00shuuuu/anomalyxl-generalization#ltaf-anomaly","mode":"rlm"}'

vf-eval timeseries_qa -m <model> \
  -a '{"dataset":"sleep_anomaly","repo_or_path":"nz00shuuuu/anomalyxl-generalization#sleep-anomaly","mode":"rlm"}'

Related

Part of the TimeRLM / AnomalyXL collection: anomalyxl-precise, anomalyxl-coarse, and this generalization set.

Provenance & licensing

Windows are spliced from real clinical recordings (long-term AF ECG; overnight polysomnography). The underlying source recordings retain their original licenses and data-use agreements — consult those before any redistribution or non-research use. This repository is a derived, re-windowed research artifact.