skyzhou06/LifeAudioMultiTurnStreamingCoT

1---2license: cc-by-4.03language:4- en5tags:6- streaming-cot7- chain-of-thought8- sft9- audio10size_categories:11- 1K<n<10K12configs:13- config_name: default14  data_files:15  - split: train16    path: data/train.parquet17  - split: eval18    path: data/eval.parquet19- config_name: high_quality20  data_files:21  - split: train22    path: data/high_quality_train.parquet23  - split: eval24    path: data/high_quality_eval.parquet25---26 27# LifeAudioMultiTurnStreamingCoT28 29**Version:** vFinal30 31A multi-turn audio dataset with streaming chain-of-thought reasoning for SFT. 13349 active training rows with audio references.32 33## Schema34 35Top-level fields: `id, split, modality, turn_type, dialogue, audio, input, streaming, target, taxonomy, quality, source, metadata`36 37- dialogue (list of turns with role/text)38- audio (path, reference, duration_sec, loadability)39- input (instruction, length_bucket)40- target (reasoning, answer, response)41- streaming (checkpoints with streaming_reasoning)42- taxonomy (category, subcategory, difficulty, intent_type)43 44### Target Format45```json46{47  "reasoning": "Natural language reasoning about the task/input...",48  "answer": "The actual task output...",49  "response": "Reasoning: ...\n\nAnswer: ..."50}51```52 53Use `target.response` for SFT training. It includes both reasoning and final answer.54 55## Quality56 57| Metric | Value |58|--------|-------|59| Active rows | 13,349 |60| Train | 10,679 |61| Eval | 2,670 |62| High quality | 13,349 |63| SFT-ready | 100.0% |64| Target grounded | 100.0% |65 66## High-Quality Configuration67 68The `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.69 70## Limitations71 72- Audio is external reference only (HF URLs). No bundled audio bytes.73- 74- Natural-language reasoning is template-generated, not LLM-written.75- Row counts reflect quality-filtered active splits suitable for direct SFT usage.76 77## Usage78 79```python80from datasets import load_dataset81 82# Load default config83ds = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT")84 85# Load high-quality subset86ds_hq = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT", "high_quality")87```88