---
dataset_info:
  features:
  - name: sentence1
    dtype: string
  - name: sentence2
    dtype: string
  - name: labels
    dtype: string
  - name: metadata
    dtype: string
  splits:
  - name: test
    num_bytes: 1201360
    num_examples: 2818
  - name: validation
    num_bytes: 825663
    num_examples: 1927
  - name: train
    num_bytes: 4681250
    num_examples: 10911
  download_size: 2642304
  dataset_size: 6708273
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
  - split: train
    path: data/train-*
license: cc-by-4.0
source_datasets: []
---

# slurp

SLURP original human transcripts and intent labels, without bundled audio.

Native real train/devel/test partitions; synthetic speech is not included.
Identical transcripts with inconsistent intents are excluded, and duplicate
transcripts stay in their highest-priority native holdout (test, dev, train).
One- and two-word transcripts are excluded as insufficient intent context.
Intent is the native scenario/action pair; inconsistent raw intent aliases
are retained in metadata. Original terms: CC BY 4.0.

Original data: [original distribution](https://github.com/pswietojanski/slurp). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/dialogue.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/dialogue.py#L24).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.