NLQTSBench-train-v2 / README.md
mrtan's picture
Add NLQTSBench-train-v2: 500 general + 190 Composite Trend training tasks
6419876 verified
|
Raw History Blame Contribute Delete
2.73 kB
metadata
license: apache-2.0
language:
  - en
task_categories:
  - question-answering
  - table-question-answering
tags:
  - time-series
  - natural-language-queries
  - benchmark
  - training-set
size_categories:
  - n<1K
pretty_name: NLQTSBench-train-v2

NLQTSBench-train-v2

Second training set for NLQTSBench. It sits next to mrtan/NLQTSBench-train, which is unchanged.

It holds 690 tasks over all 11 sub-tasks, including Composite Trend. The skill library of the Sonar-TS moso branch was distilled from it.

The tasks use the same question templates as the test set, generated with different and more widely varying parameters. No task id and no question text of this set appears in the test set. Do not use it for benchmarking.

Files

NLQTSBench-train-v2/
├── general/
│   ├── tasks.json        500 tasks, all 11 sub-tasks
│   ├── splits.json       train 300 / valid 198 / excluded 2
│   └── ts_data/<id>.csv  500 CSVs
└── composite_trend/
    ├── tasks.json        190 Composite Trend tasks
    ├── splits.json       train 120 / valid 60 / spare 10
    └── ts_data/<id>.csv  190 CSVs

The train ids feed the cold-start stream, and the valid ids select the skill library. The two excluded tasks have a degenerate interval ground truth and are not scored.

Schema

tasks.json has the same fields as the test set and v1: id, level, level_name, category, subtask, question, answer, ground_truth, eval_metric, channel, ts_data_path. There is no meta field. Each ts_data/<id>.csv is a wide table whose first column is timestamp, followed by numeric channels.

Sub-task distribution

Sub-task general composite_trend
Global Aggregation 45 0
Temporal Localization 45 0
Interval Discovery 45 0
Sliding Window 45 0
Shape Identification 45 0
Periodicity Detection 45 0
Subsequence Matching 45 0
Composite Trend 50 190
Contextual Anomaly 45 0
Causal Anomaly 45 0
Insight Synthesis 45 0
Total 500 190

Loading

from huggingface_hub import snapshot_download
snapshot_download(repo_id="mrtan/NLQTSBench-train-v2",
                  repo_type="dataset",
                  local_dir="NLQTSBench-train-v2")

import json, pandas as pd
root = "NLQTSBench-train-v2/composite_trend"
tasks = json.load(open(f"{root}/tasks.json"))
df = pd.read_csv(f"{root}/{tasks[0]['ts_data_path']}")

License

Apache-2.0, matching the test set.