Dataset Card for TURNS-2K

Dataset Summary

TURNS-2K (Turn Understanding and Recognition Natural-language Set) is a benchmark dataset for end-of-utterance (EOU) detection in conversational AI. It contains 2,000 annotated conversation turns with binary labels indicating whether each turn represents a natural ending point in the conversation. The dataset captures real-world conversational phenomena including:

• Backchannels (e.g., "uh-huh", "mm-hmm")
• Mispronunciations and self-corrections
• Code-switching and language mixing
• Colloquial variations and informal speech
• Repetitions and hesitations
• Numerical expressions
• Context shifts
• Positive and negative statements
• Interjections
• Incomplete mid-word interruptions
• Language switching between turns

Dataset Structure

Dataset({
    features: ['text', 'label'],
    num_rows: 2000
})

Label Distribution

{1: 1116, 0: 884}

Example Usage

from datasets import load_dataset

dataset = load_dataset("latishab/turns-2k")
example = dataset['train'][0]
print(example)

Dataset Creation

Data Collection and Annotation

The dataset was created through manual annotation of conversation turns, focusing on capturing diverse real-world conversational patterns. The annotations consider various aspects of natural dialogue including:

• Backchannels: Short responses that show active listening
• Speech Disfluencies: Including mispronunciations and self-corrections
• Multilingual Elements: Code-switching and language mixing patterns
• Informal Speech: Colloquialisms and conversational variations
• Turn Dynamics: Interruptions, repetitions, and context shifts

Annotations

Labels:

• 0: Non-EOU (incomplete/continuing turn)

  • Mid-sentence interruptions
  • Incomplete thoughts
  • Continuing context
  • Hesitations and fillers

• 1: EOU (complete/natural ending turn)

  • Complete statements
  • Natural conversation endings
  • Question completions
  • Proper turn transitions

Additional Information

Limitations

• The dataset may not cover all possible conversational patterns
• Annotations are based on individual judgment of turn completeness
• Regional or cultural-specific expressions may be underrepresented

License

Apache 2.0

Citation

@misc{turns2k2024,
    title = {TURNS-2K: A Benchmark Dataset for End-of-Utterance Detection},
    author = {Latisha Besariani HENDRA},
    year = {2025},
    publisher = {Hugging Face},
    howpublished = {https://huggingface.co/datasets/latishab/turns-2k},
    note = {Accessed: 2025-03-13},
}