Upload LifeAudioMultiTurnStreamingCoT v0.2: Spoken Emotional and Paralinguistic Extension

.gitattributes

@@ -58,3 +58,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+data/eval.jsonl filter=lfs diff=lfs merge=lfs -text
+data/eval_high_quality.jsonl filter=lfs diff=lfs merge=lfs -text
+data/train.jsonl filter=lfs diff=lfs merge=lfs -text
+data/train_high_quality.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,82 @@
+---
+pretty_name: LifeAudioMultiTurnStreamingCoT
+language:
+- en
+license: cc-by-nc-4.0
+version: "v0.2"
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train.parquet
+  - split: test
+    path: data/eval.parquet
+- config_name: high_quality
+  data_files:
+  - split: train
+    path: data/train_high_quality.parquet
+  - split: test
+    path: data/eval_high_quality.parquet
+task_categories:
+- text-generation
+tags:
+- audio
+- spoken-dialogue
+- multi-turn
+- streaming-reasoning
+- supervised-fine-tuning
+---
+# LifeAudioMultiTurnStreamingCoT
+
+LifeAudioMultiTurnStreamingCoT is the canonical audio/multi-turn member of the Life Streaming CoT family. It supersedes the older `LifeAudioStreamingCoT` name while preserving the v0.2 processed data.
+
+## Summary
+
+- Modality: audio
+- Turn type: multi_turn
+- Version: v0.2
+- HF repo: `skyzhou06/LifeAudioMultiTurnStreamingCoT`
+- Rows: 13355 total, 10684 train, 2671 eval
+- High-quality rows: 10588 train, 2640 eval
+- Source distribution: `{"SpokenWOZ": 9999, "MultiDialog": 3022, "SD-Eval": 334}`
+- Licenses: `{"cc": 3022, "cc-by-nc-4.0": 10333}`
+- Category distribution: `{"daily_life": 2746, "health_wellness_safe": 46, "social_communication": 3356, "travel": 7207}`
+- Length buckets: `{"short": 9146, "short_medium": 1539, "very_short": 2670}`
+
+## Audio Source Policy
+
+Per-turn audio references are valid for this multi-turn dataset. The data preserves source dialogue/session IDs, source turn IDs, order, and chunk timestamps. Raw audio is referenced through HF archive URIs, direct WAV URLs, or upstream source IDs; unrelated dialogues are not merged.
+
+## Schema
+
+Core legacy fields include `audio_chunks`, `audio_chunk_timestamps`, `transcript_chunks`, `asr_chunks`, `context_chunks`, `context`, `streaming_reasoning`, `deep_reasoning`, `answer`, `messages`, `text`, `audio_metadata`, `quality_flags`, `quality_score`, and `split`.
+
+Normalized backward-compatible fields include `modality`, `turn_type`, `taxonomy`, `input`, `streaming`, and `output`.
+
+## Taxonomy
+
+Rows include normalized `taxonomy.category`, `taxonomy.subcategory`, `taxonomy.intent_type`, and `taxonomy.difficulty`, while retaining `domain` and `task_domain` for backward compatibility.
+
+## Streaming and Deep Reasoning
+
+Streaming reasoning is generated over gradually revealed transcript/audio chunks. Deep reasoning summarizes the full observed dialogue prefix and target response.
+
+## How to use for SFT
+
+- Input: dialogue with user audio turns from `audio_chunks`/`transcript_chunks` plus `instruction`.
+- Target: `answer` or `output.answer`.
+- Optional reasoning target: `streaming_reasoning`, `deep_reasoning`, then `answer`.
+
+```python
+from datasets import load_dataset
+
+full = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT", "default")
+hq = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT", "high_quality")
+```
+
+## Limitations
+
+- Raw audio bytes are not bundled; rows keep source references and timestamp spans when available.
+- Some upstream sources use non-commercial licenses; users must respect source licenses.
+- Reasoning traces are deterministic source-grounded annotations, not human-written chain-of-thought.
+- ASR/paralinguistic noise labels are heuristic for some non-SpokenWOZ examples.

data/eval.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8335555a0dae394f2daf1e451f67c07752aee1f1a8aec401e2ba0f397e52049
+size 24634028

data/eval.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9676684191c84fecc4b1afc313d59301bb260c5fd9de0e0f6faa3a3a93dda17
+size 5163660

data/eval_high_quality.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d83ec7ef23f23f61077e28a122b4b036056444e59aec158ec247ca663c259a34
+size 24411567

data/eval_high_quality.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4889ebe47f33ea654967137ee921b68452fe8d6df40c03c1e2ade8bf22db05c1
+size 5117357

data/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5858b49e9defa4a24ce0c7122abab940f5992ede3396640fe29f130f73fb0d8
+size 97962418

data/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:483754b674664e784ab06a45a6ade7c0b4c1599629310eb0706c207aa0473791
+size 18231261

data/train_high_quality.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b30be86c64756087c4b1c8856ef89affe4279c1d9a1a4e7042a042e6984b63a3
+size 97340412

data/train_high_quality.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8489ed8fca8c2718bac353f69a0175cde7ccc416e4c68bacf0596abb81056dd2
+size 18089497

dataset_card.md

@@ -0,0 +1,82 @@
+---
+pretty_name: LifeAudioMultiTurnStreamingCoT
+language:
+- en
+license: cc-by-nc-4.0
+version: "v0.2"
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train.parquet
+  - split: test
+    path: data/eval.parquet
+- config_name: high_quality
+  data_files:
+  - split: train
+    path: data/train_high_quality.parquet
+  - split: test
+    path: data/eval_high_quality.parquet
+task_categories:
+- text-generation
+tags:
+- audio
+- spoken-dialogue
+- multi-turn
+- streaming-reasoning
+- supervised-fine-tuning
+---
+# LifeAudioMultiTurnStreamingCoT
+
+LifeAudioMultiTurnStreamingCoT is the canonical audio/multi-turn member of the Life Streaming CoT family. It supersedes the older `LifeAudioStreamingCoT` name while preserving the v0.2 processed data.
+
+## Summary
+
+- Modality: audio
+- Turn type: multi_turn
+- Version: v0.2
+- HF repo: `skyzhou06/LifeAudioMultiTurnStreamingCoT`
+- Rows: 13355 total, 10684 train, 2671 eval
+- High-quality rows: 10588 train, 2640 eval
+- Source distribution: `{"SpokenWOZ": 9999, "MultiDialog": 3022, "SD-Eval": 334}`
+- Licenses: `{"cc": 3022, "cc-by-nc-4.0": 10333}`
+- Category distribution: `{"daily_life": 2746, "health_wellness_safe": 46, "social_communication": 3356, "travel": 7207}`
+- Length buckets: `{"short": 9146, "short_medium": 1539, "very_short": 2670}`
+
+## Audio Source Policy
+
+Per-turn audio references are valid for this multi-turn dataset. The data preserves source dialogue/session IDs, source turn IDs, order, and chunk timestamps. Raw audio is referenced through HF archive URIs, direct WAV URLs, or upstream source IDs; unrelated dialogues are not merged.
+
+## Schema
+
+Core legacy fields include `audio_chunks`, `audio_chunk_timestamps`, `transcript_chunks`, `asr_chunks`, `context_chunks`, `context`, `streaming_reasoning`, `deep_reasoning`, `answer`, `messages`, `text`, `audio_metadata`, `quality_flags`, `quality_score`, and `split`.
+
+Normalized backward-compatible fields include `modality`, `turn_type`, `taxonomy`, `input`, `streaming`, and `output`.
+
+## Taxonomy
+
+Rows include normalized `taxonomy.category`, `taxonomy.subcategory`, `taxonomy.intent_type`, and `taxonomy.difficulty`, while retaining `domain` and `task_domain` for backward compatibility.
+
+## Streaming and Deep Reasoning
+
+Streaming reasoning is generated over gradually revealed transcript/audio chunks. Deep reasoning summarizes the full observed dialogue prefix and target response.
+
+## How to use for SFT
+
+- Input: dialogue with user audio turns from `audio_chunks`/`transcript_chunks` plus `instruction`.
+- Target: `answer` or `output.answer`.
+- Optional reasoning target: `streaming_reasoning`, `deep_reasoning`, then `answer`.
+
+```python
+from datasets import load_dataset
+
+full = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT", "default")
+hq = load_dataset("skyzhou06/LifeAudioMultiTurnStreamingCoT", "high_quality")
+```
+
+## Limitations
+
+- Raw audio bytes are not bundled; rows keep source references and timestamp spans when available.
+- Some upstream sources use non-commercial licenses; users must respect source licenses.
+- Reasoning traces are deterministic source-grounded annotations, not human-written chain-of-thought.
+- ASR/paralinguistic noise labels are heuristic for some non-SpokenWOZ examples.

dataset_info.json

@@ -0,0 +1,186 @@
+{
+  "dataset_name": "LifeAudioMultiTurnStreamingCoT",
+  "repo_id": "skyzhou06/LifeAudioMultiTurnStreamingCoT",
+  "version": "v0.2",
+  "created_at_utc": "2026-05-20T07:29:46Z",
+  "source_datasets": [
+    "SpokenWOZ via ssz1111/SpokenWOZ-Train-Text",
+    "SpokenWOZ via ssz1111/SpokenWOZ-Test-Text",
+    "SD-Eval via amphion/SD-Eval metadata",
+    "MultiDialog via IVLLab/MultiDialog metadata"
+  ],
+  "source_datasets_skipped": {
+    "ssz1111/SpokenWOZ-Train-Audio": "raw train/dev audio is a 10GB archive; v0.2 stores HF archive references and timestamps instead of duplicating audio",
+    "ssz1111/SpokenWOZ-Test-Audio": "test WAVs are referenced by URL; raw audio bytes are not bundled",
+    "SD-Eval test-emo raw audio": "metadata is public, but non-synthesized raw source audio requires manual upstream dataset downloads; v0.2 stores source audio ids and manual source paths",
+    "MultiDialog raw audio bytes": "HF audio archives are very large; v0.2 stores archive-member references instead of bundling audio",
+    "pirxus/spokenwoz-whisper": "third-party Whisper parquet mirror skipped to keep v0.2 grounded in official/public source metadata"
+  },
+  "source_access_notes": "SpokenWOZ text/timing, SD-Eval metadata, and MultiDialog metadata were loaded from public Hugging Face dataset repos. Raw audio is referenced, not redistributed.",
+  "audio_availability": {
+    "raw_audio_bundled": false,
+    "audio_reference_strategy": "HF archive URIs, direct HF WAV URLs, or manual source ids when upstream audio requires separate download",
+    "has_audio_percentage": 99.63,
+    "has_asr_percentage": 74.87
+  },
+  "paralinguistic_cue_availability": {
+    "has_paralinguistic_cues_percentage": 29.31,
+    "has_environmental_cues_percentage": 2.13
+  },
+  "total_rows": 13355,
+  "train_rows": 10684,
+  "eval_rows": 2671,
+  "high_quality_rows": 13228,
+  "train_high_quality_rows": 10588,
+  "eval_high_quality_rows": 2640,
+  "rows_by_domain": {
+    "spoken_task_oriented_assistant": 9999,
+    "spoken_daily_dialogue": 3306,
+    "spoken_emotional_support": 50
+  },
+  "rows_by_source": {
+    "SpokenWOZ": 9999,
+    "MultiDialog": 3022,
+    "SD-Eval": 334
+  },
+  "task_domains": {
+    "restaurant": 2348,
+    "taxi": 974,
+    "daily_dialogue": 3306,
+    "hotel": 2363,
+    "attraction": 1232,
+    "train": 2638,
+    "hospital": 46,
+    "unknown": 271,
+    "police": 65,
+    "emotional_support": 50,
+    "general": 62
+  },
+  "emotion_label_distribution": {
+    "null": 10283,
+    "happy": 426,
+    "curious": 1370,
+    "excited": 539,
+    "frustrated": 25,
+    "neutral": 611,
+    "worried": 13,
+    "sad": 59,
+    "angry": 29
+  },
+  "schema_fields": [
+    "id",
+    "domain",
+    "task_domain",
+    "source_dataset",
+    "source_dialogue_id",
+    "source_turn_ids",
+    "source_audio_id",
+    "instruction",
+    "audio_path",
+    "audio_url",
+    "audio_chunks",
+    "audio_chunk_timestamps",
+    "transcript_chunks",
+    "asr_chunks",
+    "context_chunks",
+    "context",
+    "streaming_reasoning",
+    "deep_reasoning",
+    "answer",
+    "response",
+    "messages",
+    "text",
+    "num_chunks",
+    "language",
+    "split",
+    "generation_method",
+    "version",
+    "reasoning_policy",
+    "chunking_method",
+    "chunk_labels",
+    "skip_chunks",
+    "skip_reasons",
+    "speech_event_labels",
+    "paralinguistic_cues",
+    "environmental_cues",
+    "emotion_label",
+    "emotion_intensity",
+    "speaker_state",
+    "paralinguistic_summary",
+    "source_task_type",
+    "source_label_schema",
+    "source_license",
+    "modality_available",
+    "asr_noise_flags",
+    "dialogue_state",
+    "audio_metadata",
+    "quality_flags",
+    "quality_score",
+    "is_high_quality",
+    "rejected_reason"
+  ],
+  "generation_method": "source_grounded_rule_based_v0.2_spoken_emotional_extension",
+  "reasoning_policy": "selective_concise",
+  "chunking_method": "spoken_turn_or_semantic_segment_v0.2",
+  "average_quality_score": 0.9918,
+  "high_quality_percentage": 99.05,
+  "quality_flags_distribution": {
+    "weak_context": 333,
+    "repeated_context_chunks": 399,
+    "missing_audio_reference": 50,
+    "missing_audio_timestamps": 50,
+    "too_many_skips": 161,
+    "awkward_answer": 30
+  },
+  "limitations": [
+    "v0.2 is speech-text grounded, not speech-to-speech.",
+    "Raw audio files are not bundled; rows keep source references and timestamp spans when available.",
+    "Streaming and deep reasoning fields are generated by deterministic source-grounded rules.",
+    "ASR noise labels are simple heuristics; most non-SpokenWOZ examples are transcript-only.",
+    "SD-Eval emotion/environment labels are source metadata, while paralinguistic summaries are normalized from those labels.",
+    "The dataset inherits upstream source licenses, including CC BY-NC 4.0 for SpokenWOZ/SD-Eval."
+  ],
+  "build_source_stats": {
+    "source_dialogues_loaded": 5700,
+    "spokenwoz_used": 9999,
+    "test-emo_metadata_rows": 1289,
+    "test-env_metadata_rows": 690,
+    "sdeval_used": 334,
+    "valid_freq_metadata_rows": 8513,
+    "valid_rare_metadata_rows": 9556,
+    "test_freq_metadata_rows": 9810,
+    "test_rare_metadata_rows": 8331,
+    "multidialog_used": 3022
+  },
+  "old_dataset_name": "LifeAudioStreamingCoT",
+  "old_repo_id": "skyzhou06/LifeAudioStreamingCoT",
+  "hf_repo_url": "https://huggingface.co/datasets/skyzhou06/LifeAudioMultiTurnStreamingCoT",
+  "modality": "audio",
+  "turn_type": "multi_turn",
+  "source_licenses": {
+    "cc": 3022,
+    "cc-by-nc-4.0": 10333
+  },
+  "category_distribution": {
+    "daily_life": 2746,
+    "health_wellness_safe": 46,
+    "social_communication": 3356,
+    "travel": 7207
+  },
+  "length_bucket_distribution": {
+    "short": 9146,
+    "short_medium": 1539,
+    "very_short": 2670
+  },
+  "timestamp_quality_distribution": {
+    "referenced_or_source_timed": 13355
+  },
+  "sft_readiness": {
+    "status": "PASS",
+    "blockers": []
+  },
+  "publishability_gate": {
+    "status": "PASS",
+    "blockers": []
+  }
+}