NLQTSBench-train

Training-split companion to mrtan/NLQTSBench. 220 natural-language-query / time-series tasks sampled from the same candidate pool as the NLQTSBench test set, then filtered to be disjoint by id from the test split.

Intended use: cold-start training of methodology skills for the Sonar-TS framework. Do not use this split for benchmarking — it overlaps the test set's task templates even though no individual task id appears in both.

Files

NLQTSBench-train/
├── tasks.json          220 task records (one JSON list)
└── ts_data/
    └── <id>.csv        220 raw time-series CSVs (one per task)

Schema

Each entry in tasks.json has these fields:

Field	Type	Description
id	str	Unique task id, e.g. L1_T1_Global_Aggregation_00014.
level	int (1–4)	Reasoning level (paper Table 2).
level_name	str	Human-readable level name (Basic Operations, …).
category	str	Higher-level category (Atomic Retrieval, Pattern Recognition, …).
subtask	str	One of 10 sub-task labels (see counts below — Composite Trend is absent).
question	str	Natural-language query, includes the expected answer format.
answer	str	Ground-truth answer as a string (display form).
ground_truth	varies	Ground-truth value in its native Python type (number / list / dict).
eval_metric	str	One of rel_acc, hit, iou, set_f1, report.
channel	str / list	Channel name(s) referenced by the question.
ts_data_path	str	Relative path to the matching CSV under ts_data/.
meta	dict	Generation metadata (template args, source dataset).

Each ts_data/<id>.csv is a wide table whose first column is timestamp and remaining columns are numeric channels.

Sub-task distribution

Sub-task	N
Sliding Window	38
Interval Discovery	33
Temporal Localization	31
Global Aggregation	28
Causal Anomaly	18
Subsequence Matching	16
Contextual Anomaly	14
Insight Synthesis	14
Periodicity Detection	14
Shape Identification	14
Composite Trend	0
Total	220

Composite Trend is intentionally absent: its candidate pool was fully consumed by the test split with no disjoint tasks left over.

Disjointness with the test split

Tasks were sampled with a distinct seed and then any id that also appears in mrtan/NLQTSBench was dropped. The remaining 220 task ids do not overlap with the test set.

Loading

from huggingface_hub import snapshot_download
snapshot_download(repo_id="mrtan/NLQTSBench-train",
                  repo_type="dataset",
                  local_dir="NLQTSBench-train")

import json, pandas as pd
tasks = json.load(open("NLQTSBench-train/tasks.json"))
df    = pd.read_csv(f"NLQTSBench-train/{tasks[0]['ts_data_path']}")

For the Sonar-TS framework, just run:

python -m cold_start.download_train_data

which pulls this dataset and builds per-task SQLite + SAX feature tables locally under cold_start/train_data/.

License

Apache-2.0, matching the test-split licence.