README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - audio
  - function-calling
  - voxtral
  - multimodal
  - speech-processing
  - embeddings
pretty_name: Audio Function Calling (Voxtral BF16 Features)
dataset_info:
  features:
    - name: sample_id
      dtype: string
    - name: persona
      dtype: string
    - name: tools_json
      dtype: string
    - name: conversation
      dtype: string

Audio Function Calling Dataset (vxltxrllc/audio-function-calling-v1)

Multimodal dataset containing conversational function-calling turns with pre-extracted Voxtral Mini (vxltxrllc/voxtral-mini-audio-extractor) latent audio features in bfloat16.

Overview

• Audio Turns: 703 user turns processed across 137 multi-turn conversations (~85.7 minutes of effective speech audio).
• Feature Format: Unpadded sequence-trimmed bfloat16 tensors with shape [T_frames, 3072] stored in .safetensors shards (12.5 Hz frame rate).
• Manifest: Compact Parquet manifest (data/train.parquet, ~523 KB) linking text conversations and function calls directly to latent feature shards (features/shard_*.safetensors).

Dataset Structure

vxltxrllc/audio-function-calling-v1/
├── README.md
├── data/
│   └── train.parquet              # Main conversational manifest
└── features/
    ├── shard_00000.safetensors    # BF16 audio feature embeddings [T, 3072]
    └── shard_00001.safetensors