README.md

metadata

language:
  - en
  - hi
pretty_name: DECIDE — 20 Public Samples
task_categories:
  - audio-classification
  - question-answering
size_categories:
  - n<1K
tags:
  - audio
  - speech
  - voice-agents
  - decision-making
  - multiple-choice
  - benchmark-preview
configs:
  - config_name: default
    data_files:
      - split: preview
        path: data/preview-00000-of-00001.parquet

DECIDE: 20 public samples

Decision Evaluation in Calls: Intent, Dialogue and Emotion

Website · GitHub · Full benchmark evaluation

This public preview contains exactly 20 labelled decision questions, using six audio clips already published on the DECIDE website. A clip can support multiple different questions. The examples span 14 task types across all six decision families. Audio comes from production voice-agent calls supplied by Bolna.

The full DECIDE v1.0 benchmark contains 4,867 questions over 1,081 clips and 32 tasks and remains private. This repository is a small, deliberately selected demonstration of question types, not the full benchmark or a representative test set. Use the hosted evaluator to obtain full-benchmark results.

Contents

Decision family	Rows
Business and conversation	7
Graph routing	4
Language switching	1
Tool decisions	1
Turn taking	3
Voice properties	4
Total	20

Examples include lead qualification, caller availability and callback requests, information delivery, caller responses, graph routing and next-node selection, external-tool selection, next-turn prediction, caller emotion, speaker count, audio disturbance and language switching.

Each row contains:

• audio: a playable mono 16 kHz PCM16 WAV clip, embedded in the Parquet file.
• question, context, options: the decision question, permitted context and allowed option IDs with descriptions.
• answer: the existing human reference label from the frozen DECIDE v1.0 data.
• task, family: the task identifier and decision family.
• id, clip_id: public sample identifiers. Repeated clip_id values mean the same audio is being used for different questions.
• current_node: the node at the clip start for graph questions; empty for other questions. The graph specification is included in that row's context.
• duration_seconds, duration_bucket: audio duration and duration group.

Load the preview

from datasets import Audio, load_dataset

samples = load_dataset("tazolabs/DECIDE-samples", split="preview")
assert len(samples) == 20

# Read encoded WAV bytes without installing an audio decoder.
samples = samples.cast_column("audio", Audio(decode=False))
row = samples[0]
print(row["question"])
print(row["options"])
print(row["answer"])
wav_bytes = row["audio"]["bytes"]

The Hugging Face dataset viewer provides an audio player for each row. For local decoded-audio playback, use the audio dependencies supported by your version of datasets.

Interpretation

The task is to choose one allowed decision at the end of the supplied clip. Use only the audio and permitted context. Future conversation is not provided. agent means the AI assistant; caller means the human/user. Graphs and tools are reconstructed benchmark specifications, not production execution logs. Choosing a tool does not execute it.

Reference answers are copied unchanged from existing human annotations. Context and question wording are shortened for public inspection; full-call histories, private source identifiers, reviewer notes and model predictions are excluded. Consequently, preview scores must not be reported as full DECIDE benchmark scores.

Only these 20 question rows are published here. No new call recordings beyond the six existing public website previews are released. The code repository's license does not grant a license to the underlying call recordings; no separate open-data license is specified for this preview.

Attribution

DECIDE is created by Tazo Labs using production call recordings supplied by Bolna. Authors: Harsh Pare, Naman Jain, Sachin Kumar Singh, Jayakrishna Sahit, Siddharth Batra, Sanket Bhat and Surya Vamsee Ayyagari. Paper/arXiv link: forthcoming. See the website for updates.