README.md

---
pretty_name: LifeMultiTurnStreamingCoT
language:
- en
license: apache-2.0
version: "v0.4"
task_categories:
- text-generation
tags:
- streaming-reasoning
- rule-based-reasoning
- explicit-reasoning
- multi-turn-dialogue
- life-assistant
- supervised-fine-tuning
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: eval
    path: eval.jsonl
  - split: high_quality
    path: high_quality.jsonl
---
# LifeMultiTurnStreamingCoT

Version: v0.4 — Explicit Streaming Reasoning Release

LifeMultiTurnStreamingCoT is a text-to-text multi-turn life-domain dataset. Each row uses previous user-assistant dialogue turns as input, deterministic turn-level streaming state tracking as intermediate supervision, a compact final-state-based deep reasoning summary, and the next assistant turn as the answer.

This v0.4 release keeps the real DailyDialog, MultiWOZ 2.2, and Taskmaster sources while adding explicit turn-aligned streaming reasoning. No LLM calls are used for construction. The reasoning traces are deterministic rule-based annotations over visible dialogue prefixes and repaired state deltas.

## Current Build Statistics

- Version: v0.4 — Explicit Streaming Reasoning Release
- Total rows: 30000
- Train rows: 24211
- Eval rows: 5789
- High-quality split rows: 17723
- High-quality train rows: 14339
- High-quality eval rows: 3384
- Average input turns: 9.705
- Average streaming chunks: 9.705
- Source distribution: {"DailyDialog": 10000, "MultiWOZ": 10000, "Taskmaster": 10000}
- Domain category distribution: {"customer_service": 204, "education_career": 2288, "finance_business": 180, "food_dining": 1262, "general_daily_life": 2881, "health_wellness": 221, "home_services": 226, "hospitality_lodging": 404, "personal_schedule": 688, "shopping_retail": 527, "social_relationship": 1984, "technology_support": 193, "travel_transportation": 18942}
- Intent category distribution: {"booking_or_reservation": 15758, "confirmation_clarification": 2710, "customer_support": 1241, "emotional_support": 232, "information_request": 6242, "instruction_following": 26, "negotiation_decision": 139, "planning_coordination": 1460, "problem_solving": 440, "recommendation": 739, "small_talk": 1013}
- Scenario category distribution: {"attraction_search": 455, "banking_support": 210, "customer_complaint": 218, "family_conversation": 600, "flight_booking": 9926, "food_ordering": 18, "friend_conversation": 456, "general_conversation": 3288, "home_repair": 101, "hotel_booking": 1557, "hotel_search": 184, "insurance_support": 95, "job_interview": 104, "medical_assistance": 291, "movie_ticketing": 218, "music_search": 141, "restaurant_booking": 1890, "restaurant_search": 819, "schedule_planning": 646, "school_life": 853, "shopping_assistance": 393, "taxi_booking": 1109, "technical_support": 453, "train_booking": 4317, "travel_planning": 253, "workplace_conversation": 1405}
- Taxonomy confidence distribution: {"high": 20506, "low": 2881, "medium": 6613}
- Quality tier distribution: {"bronze": 7505, "drop": 4772, "gold": 12479, "silver": 5244}
- Safety category distribution: {"safe": 28743, "sensitive": 1257}
- Target answer quality distribution: {"drop": 3924, "strong": 15129, "usable": 6819, "weak": 4128}
- Streaming reasoning quality distribution: {"strong": 14767, "usable": 15233}
- Streaming reasoning confidence distribution: {"high": 201954, "medium": 89194}
- Slot repair rows: 15233
- Suppressed slot reasons: {"generic_number_misread_as_budget": 36309, "percent_misread_as_budget": 13}
- Unknown/other taxonomy ratio: 0.0
- Category distribution: {"daily_dialogue": 10000, "task_oriented_dialogue": 20000}

## Sources

- DailyDialog: daily multi-turn dialogue.
- MultiWOZ 2.2: multi-domain task-oriented dialogue.
- Taskmaster: real task-oriented dialogue from Taskmaster conversations.

## Schema

Rows contain `id`, `source_dataset`, `source_id`, `dialogue_id`, `domain`, `task_type`, `dialogue_history`, `streaming_chunks`, top-level `streaming_reasoning`, `deep_reasoning`, `answer`, `metadata`, `quality_flags`, `quality_score`, `is_high_quality`, and `split`.

`streaming_chunks[i]` includes the original chunk fields plus `state_before`, `state_update`, `state_after`, `state_delta`, `reasoning`, `reasoning_type`, and `reasoning_confidence`. The top-level `streaming_reasoning[i]` list is aligned with `streaming_chunks[i]` by `chunk_id` and `turn_id`. The final schema remains unified across sources; source-specific details such as source, category, domain/services, scenario, original split, and raw file are kept in `metadata`.

## Splits

- `train`: training rows after deterministic dialogue-level split.
- `eval`: evaluation rows after deterministic dialogue-level split.
- `high_quality`: rows from train+eval where `is_high_quality` is true, `quality_tier` is gold/silver, safety is safe, and explicit streaming reasoning is strong/usable. Each high-quality row keeps `metadata.original_release_split`.

## Version History

### v0.4 — Explicit Streaming Reasoning Release

- Added top-level `streaming_reasoning` aligned with `streaming_chunks`.
- Added per-chunk `state_delta`, `reasoning`, `reasoning_type`, and `reasoning_confidence`.
- Added deterministic slot repair/suppression for unsupported visible-prefix slot evidence, including budget numeric misreads.
- Added `metadata.streaming_reasoning_method`, `metadata.has_explicit_streaming_reasoning`, `metadata.streaming_reasoning_quality`, `metadata.slot_repair_applied`, and `metadata.suppressed_slots`.
- Added explicit reasoning quality flags and a `high_quality` split.

### v0.3 — Safety and Grounding Quality Release

- Added safety/content filtering for training suitability.
- Added grounded slot extraction checks.
- Added target answer usefulness checks.
- Added taxonomy confidence and evidence fields.
- Added quality tiers: gold, silver, bronze, and drop.
- Updated `is_high_quality` to use quality tiers.
- Removed or downgraded unsafe, ungrounded, malformed, or low-usefulness samples.

### v0.2 — Taxonomy-Aware Quality Release

- Added `domain_category`, `intent_category`, and `scenario_category`.
- Added deterministic source-aware taxonomy rules.
- Improved category-aware quality thresholds.
- Reduced false penalties for long task-oriented dialogues.

### v0.1 — Initial Real-Source Release

- Added real DailyDialog, MultiWOZ 2.2, and Taskmaster data.
- Built a balanced multi-source dataset with rule-based streaming/deep reasoning.

## Category Taxonomy

Each sample includes a coarse `metadata.category` and three additional taxonomy fields:

- `metadata.domain_category`: broad topic/domain such as travel, dining, lodging, entertainment, health, education, work, shopping, or general daily life.
- `metadata.intent_category`: interaction intent such as information request, recommendation, booking, planning, customer support, small talk, or emotional support.
- `metadata.scenario_category`: more specific scenario such as restaurant booking, hotel search, taxi booking, train booking, food ordering, movie ticketing, schedule planning, or workplace conversation.

The taxonomy is deterministic and source-aware. MultiWOZ uses domain/service annotations, Taskmaster uses scenario/file metadata when available, and DailyDialog uses lightweight keyword and dialogue-pattern rules. DailyDialog rows may have low taxonomy confidence when only weak keyword evidence is available.

## Quality Metadata

Each row includes additional metadata fields:

- `metadata.safety_category`
- `metadata.safety_flags`
- `metadata.is_safe_for_training`
- `metadata.grounding_flags`
- `metadata.slot_grounding_score`
- `metadata.has_grounding_issue`
- `metadata.target_answer_flags`
- `metadata.target_answer_quality`
- `metadata.taxonomy_confidence`
- `metadata.taxonomy_evidence`
- `metadata.quality_tier`
- `metadata.streaming_reasoning_method`
- `metadata.has_explicit_streaming_reasoning`
- `metadata.streaming_reasoning_quality`
- `metadata.slot_repair_applied`
- `metadata.suppressed_slots`

Recommended default training filter:

```python
row["is_high_quality"] is True
and row["metadata"]["is_safe_for_training"] is True
and row["metadata"]["quality_tier"] in ["gold", "silver"]
and row["metadata"]["has_explicit_streaming_reasoning"] is True
and row["metadata"]["streaming_reasoning_quality"] in ["strong", "usable"]
```

## Reasoning

Streaming reasoning is generated by deterministic rule-based state tracking over turn-level chunks. v0.4 explicit reasoning is generated only from visible prefix state, source metadata, repaired state deltas, and safety/taxonomy metadata. It does not call an LLM and does not rewrite the answer. DailyDialog rows focus on daily context and continuity. MultiWOZ and Taskmaster rows use task-oriented state changes, missing-slot status, and next-step policy. Deep reasoning is a compact global summary from the final tracked state, dialogue history, and target answer.

## Quality Filtering

The quality checks are category-aware and include safety, grounding, target usefulness, explicit streaming reasoning quality, and tiering. Long task-oriented conversations are no longer penalized in the same way as short daily dialogues. Some heuristic checks, including weak final state and premature response detection, are kept as diagnostic warnings rather than hard filters.

`quality_flags` and `metadata.quality_checks` support filtering by real-source status, multi-turn context, non-empty reasoning, placeholder detection, category-aware length checks, malformed-row checks, repetition checks, grounding checks, safety checks, target-answer checks, explicit reasoning checks, and role alternation. Raw external data is not committed to git; processed train/eval/high_quality files are intended for upload to `skyzhou06/LifeMultiTurnStreamingCoT`.

## Leakage Control

Train/eval splitting is performed by `dialogue_id`, so prefix samples from the same dialogue do not appear in both splits.