metadata
license: cc-by-4.0
language:
- en
tags:
- streaming-cot
- chain-of-thought
- sft
- audio
size_categories:
- 1K<n<10K
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
LifeAudioSingleTurnStreamingCoT
Version: vFinal
A single-turn audio dataset with streaming chain-of-thought reasoning for SFT. 822 active training rows with audio references.
Schema
Top-level fields: id, split, modality, turn_type, audio, input, timestamps, streaming, target, taxonomy, quality, source, metadata
- audio (path, reference, duration_sec, loadability)
- input (text, instruction, length_bucket)
- target (reasoning, answer, response)
- streaming (checkpoints with streaming_reasoning)
- taxonomy (category, subcategory, difficulty, intent_type)
Target Format
{
"reasoning": "Natural language reasoning about the task/input...",
"answer": "The actual task output...",
"response": "Reasoning: ...\n\nAnswer: ..."
}
Use target.response for SFT training. It includes both reasoning and final answer.
Quality
| Metric | Value |
|---|---|
| Active rows | 822 |
| Train | 685 |
| Eval | 137 |
| High quality | 822 |
| SFT-ready | 100.0% |
| Target grounded | 100.0% |
High-Quality Configuration
The high_quality config contains a filtered subset of default rows where quality.sft_ready = true and quality.is_high_quality = true. It is not additional unique data.
Limitations
- Audio is external reference only (HF URLs). No bundled audio bytes.
- Natural-language reasoning is template-generated, not LLM-written.
- Row counts reflect quality-filtered active splits suitable for direct SFT usage.
Usage
from datasets import load_dataset
# Load default config
ds = load_dataset("skyzhou06/LifeAudioSingleTurnStreamingCoT")
# Load high-quality subset
ds_hq = load_dataset("skyzhou06/LifeAudioSingleTurnStreamingCoT", "high_quality")