Datasets:

skyzhou06
/

LifeSpeechQAStreamingCoT

license: cc-by-4.0
language:
  - multilingual
tags:
  - streaming-cot
  - chain-of-thought
  - speech
  - qa
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
      - split: eval
        path: data/eval.parquet
  - config_name: high_quality
    data_files:
      - split: train
        path: data/high_quality_train.parquet
      - split: eval
        path: data/high_quality_eval.parquet

LifeSpeechQAStreamingCoT — Source-Expanded Release

Version: vNext-source-expanded

Overview

Multi-source release. This release adds source diversity beyond the original FLEURS seed data.

Source Distribution

FLEURS: 4091 rows ()
VoxPopuli: 4321 rows ()
MultilingualLibriSpeech: ❌

Stats

Metric	Value
Active rows	8,412
Train/Eval	6,944/1,468
High quality	8,412
SFT-ready	100.0%
Target grounded	99.5%

Schema

Top-level fields: id, split, task_type, turn_type, input_modality, output_modality, audio, input, timestamps, streaming, target, quality, source, metadata

Use target.response for SFT training.

Limitations

Audio is external reference only (no bundled bytes)
QA is FLEURS + VoxPopuli only; additional speech QA sources limited
Reasoning is template-generated, not LLM-written

Usage

from datasets import load_dataset
ds = load_dataset("skyzhou06/LifeSpeechQAStreamingCoT")
ds_hq = load_dataset("skyzhou06/LifeSpeechQAStreamingCoT", "high_quality")