| --- |
| pretty_name: LifeStreamingCoT |
| language: |
| - en |
| license: apache-2.0 |
| version: "v0.4.1" |
| 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: |
| - streaming-reasoning |
| - selective-reasoning |
| - quality-refined |
| - supervised-fine-tuning |
| - sft |
| - dialogue |
| - task-oriented-dialogue |
| - life-assistant |
| - streamingthinker |
| size_categories: |
| - 1K<n<10K |
| --- |
| # LifeStreamingCoT |
|
|
| Current version: v0.4.1: Loading Config and High-Quality Subset Patch |
|
|
| LifeStreamingCoT is a text-only, life-scenario adaptation of StreamingCoT-style data for StreamingThinker-style supervised fine-tuning. It keeps compatibility with earlier LifeStreamingCoT schemas while adding quality metadata and high-quality subset files. |
|
|
| ## Version 0.4: Quality Refinement |
|
|
| v0.3 introduced selective concise streaming reasoning, semantic chunk splitting, skip labels, and chunk-level metadata. v0.4 improved quality by fixing keyword-stitching in emotional support examples, reducing daily-dialogue intent mistakes, replacing vague task-oriented updates, reducing fragment chunks, and adding `quality_score` plus `is_high_quality`. |
|
|
| v0.4 also provides high-quality subset files: |
|
|
| - `data/train_high_quality.jsonl` |
| - `data/eval_high_quality.jsonl` |
| - `data/train_high_quality.parquet` |
| - `data/eval_high_quality.parquet` |
|
|
|
|
| ## Version 0.4.1: Loading Config and High-Quality Subset Patch |
|
|
| v0.4.1 is a patch over v0.4. It fixes Hugging Face loading behavior by adding explicit dataset card configs. The default config now points only to the full dataset files, while the `high_quality` config points only to the stricter high-quality subset files. |
|
|
| ```python |
| from datasets import load_dataset |
| |
| full = load_dataset("skyzhou06/LifeStreamingCoT", "default") |
| hq = load_dataset("skyzhou06/LifeStreamingCoT", "high_quality") |
| ``` |
|
|
| Expected split sizes for v0.4.1: |
|
|
| - `default/train`: 7457 |
| - `default/test`: 1865 |
| - `high_quality/train`: 2570 |
| - `high_quality/test`: 634 |
|
|
| The high-quality subset excludes copied-source responses, awkward answer templates, keyword-stitching, repeated context chunks, weak candidates, and severe quality flags. |
|
|
| ## Version History |
|
|
| | Version | Summary | |
| | --- | --- | |
| | v0.1 | Schema-complete source-grounded baseline | |
| | v0.2 | More specific rule-based reasoning and quality flags | |
| | v0.3 | Selective concise reasoning, skip labels, semantic chunking | |
| | v0.4 | Quality refinement, quality scores, high-quality subset | |
| | v0.4.1 | HF loading config fix, stricter high-quality filtering | |
|
|
| ## Recommended Usage |
|
|
| Full dataset: |
|
|
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("skyzhou06/LifeStreamingCoT") |
| ``` |
|
|
| Quality filtering: |
|
|
| ```python |
| clean = ds.filter(lambda x: x["is_high_quality"] and x["quality_score"] >= 0.85) |
| ``` |
|
|
| Removing flagged data: |
|
|
| ```python |
| clean = ds.filter(lambda x: len(x["quality_flags"]) == 0) |
| ``` |
|
|
| ## Schema |
|
|
| - `id` |
| - `domain` |
| - `source_dataset` |
| - `instruction` |
| - `context` |
| - `context_chunks` |
| - `streaming_reasoning` |
| - `deep_reasoning` |
| - `answer` |
| - `response` |
| - `messages` |
| - `text` |
| - `num_chunks` |
| - `language` |
| - `split` |
| - `generation_method` |
| - `quality_flags` |
| - `version` |
| - `reasoning_policy` |
| - `chunking_method` |
| - `chunk_labels` |
| - `skip_chunks` |
| - `skip_reasons` |
| - `reasoning_token_budget` |
| - `original_num_chunks` |
| - `chunk_split_count` |
| - `quality_score` |
| - `is_high_quality` |
| - `refinement_method` |
| - `llm_augmented` |
| - `llm_augmentation_model` |
| - `rejected_reason` |
| - `state_tracking_confidence` |
|
|
| ## Source Datasets |
|
|
| Used sources: |
|
|
| - `b-mc2/wikihow_lists`: 622 rows, domain `how_to_guidance` |
| - `pietrolesci/multiwoz_all_versions`: 2987 rows, domain `task_oriented_assistant` |
| - `pixelsandpointers/better_daily_dialog`: 3713 rows, domain `daily_dialogue` |
| - `pixelsandpointers/empathetic_dialogues_for_lm`: 2000 rows, domain `emotional_support` |
|
|
| Skipped sources: |
|
|
| - None |
|
|
| ## Splits |
|
|
| - Train: 7457 |
| - Eval: 1865 |
| - Total: 9322 |
| - High-quality train: 2570 |
| - High-quality eval: 634 |
|
|
| ## Statistics |
|
|
| - Average chunks: 6.45 |
| - Average chunk length: 8.95 |
| - Average streaming reasoning words: 28.50 |
| - Average deep reasoning words: 16.48 |
| - Average quality score: 0.989 |
| - High-quality percentage: 96.70% |
| - Skip chunk ratio: 0.1119 |
| - LLM augmented rows: 0 |
|
|
| ## Quality Flags |
|
|
| - `closing_mishandled`: 7 |
| - `copied_source_response`: 469 |
| - `excessive_chunking`: 174 |
| - `long_deep_reasoning`: 2 |
| - `merged_fragments`: 5797 |
| - `too_many_skips`: 2 |
| - `weak_context`: 10 |
|
|
| ## Example |
|
|
| ```json |
| { |
| "id": "life_task_oriented_assistant_000001", |
| "domain": "task_oriented_assistant", |
| "source_dataset": "pietrolesci/multiwoz_all_versions", |
| "instruction": "Help the user complete a real-life task based on gradually revealed information.", |
| "context": "Chunk 1: I'm looking for a restaurant called the gandhi.\nChunk 2: Can you book it for 8 people on Sunday at 14:45?\nChunk 3: I would also like to find somewhere fun in the middle of town to go.\nChunk 4: Sorry, I actually wanted somewhere in the west.\nChunk 5: Sure, please provide me with the postcode and address.\nChunk 6: Thank you for your help, that is all I need today.", |
| "context_chunks": [ |
| "I'm looking for a restaurant called the gandhi.", |
| "Can you book it for 8 people on Sunday at 14:45?", |
| "I would also like to find somewhere fun in the middle of town to go.", |
| "Sorry, I actually wanted somewhere in the west.", |
| "Sure, please provide me with the postcode and address.", |
| "Thank you for your help, that is all I need today." |
| ], |
| "streaming_reasoning": "C1 domain=restaurant; goal=search. C2 goal=booking; time+=8, 14:45; day+=sunday; party_size+=8 people. C3 destination+=find somewhere fun in the middle of town. C4 area+=west. C5 goal=request_info; requested_info+=address, postcode. C6 [SKIP: closing_only].", |
| "deep_reasoning": "Need domain=restaurant; goal=search, booking, request_info; area=west; party_size=8 people; destination=find somewhere fun in the middle of town; requested_info=address, postcode; when=sunday, 8, 14:45; closing_detected.", |
| "answer": "You're welcome. Glad I could help; have a great day.", |
| "response": "Streaming reasoning: C1 domain=restaurant; goal=search. C2 goal=booking; time+=8, 14:45; day+=sunday; party_size+=8 people. C3 destination+=find somewhere fun in the middle of town. C4 area+=west. C5 goal=request_info; requested_info+=address, postcode. C6 [SKIP: closing_only].\n\nDeep reasoning: Need domain=restaurant; goal=search, booking, request_info; area=west; party_size=8 people; destination=find somewhere fun in the middle of town; requested_info=address, postcode; when=sunday, 8, 14:45; closing_detected.\n\nAnswer: You're welcome. Glad I could help; have a great day.", |
| "messages": [ |
| { |
| "role": "user", |
| "content": "Instruction: Help the user complete a real-life task based on gradually revealed information.\n\nContext:\nChunk 1: I'm looking for a restaurant called the gandhi.\nChunk 2: Can you book it for 8 people on Sunday at 14:45?\nChunk 3: I would also like to find somewhere fun in the middle of town to go.\nChunk 4: Sorry, I actually wanted somewhere in the west.\nChunk 5: Sure, please provide me with the postcode and address.\nChunk 6: Thank you for your help, that is all I need today." |
| }, |
| { |
| "role": "assistant", |
| "content": "Streaming reasoning: C1 domain=restaurant; goal=search. C2 goal=booking; time+=8, 14:45; day+=sunday; party_size+=8 people. C3 destination+=find somewhere fun in the middle of town. C4 area+=west. C5 goal=request_info; requested_info+=address, postcode. C6 [SKIP: closing_only].\n\nDeep reasoning: Need domain=restaurant; goal=search, booking, request_info; area=west; party_size=8 people; destination=find somewhere fun in the middle of town; requested_info=address, postcode; when=sunday, 8, 14:45; closing_detected.\n\nAnswer: You're welcome. Glad I could help; have a great day." |
| } |
| ], |
| "text": "<|user|>\nInstruction: Help the user complete a real-life task based on gradually revealed information.\n\nContext:\nChunk 1: I'm looking for a restaurant called the gandhi.\nChunk 2: Can you book it for 8 people on Sunday at 14:45?\nChunk 3: I would also like to find somewhere fun in the middle of town to go.\nChunk 4: Sorry, I actually wanted somewhere in the west.\nChunk 5: Sure, please provide me with the postcode and address.\nChunk 6: Thank you for your help, that is all I need today.\n<|assistant|>\nStreaming reasoning: C1 domain=restaurant; goal=search. C2 goal=booking; time+=8, 14:45; day+=sunday; party_size+=8 people. C3 destination+=find somewhere fun in the middle of town. C4 area+=west. C5 goal=request_info; requested_info+=address, postcode. C6 [SKIP: closing_only].\n\nDeep reasoning: Need domain=restaurant; goal=search, booking, request_info; area=west; party_size=8 people; destination=find somewhere fun in the middle of town; requested_info=address, postcode; when=sunday, 8, 14:45; closing_detected.\n\nAnswer: You're welcome. Glad I could help; have a great day.", |
| "num_chunks": 6, |
| "language": "en", |
| "split": "train", |
| "generation_method": "source_grounded_rule_based_v0.4_quality_refined", |
| "quality_flags": [], |
| "version": "v0.4", |
| "reasoning_policy": "selective_concise", |
| "chunking_method": "semantic_sentence_split_v0.4_refined", |
| "chunk_labels": [ |
| "reason", |
| "reason", |
| "reason", |
| "reason", |
| "reason", |
| "skip" |
| ], |
| "skip_chunks": [ |
| 6 |
| ], |
| "skip_reasons": { |
| "6": "closing_only" |
| }, |
| "reasoning_token_budget": { |
| "streaming_reasoning_max_words_per_chunk": 18, |
| "deep_reasoning_max_words": 45, |
| "answer_max_sentences": 3 |
| }, |
| "original_num_chunks": 6, |
| "chunk_split_count": 0, |
| "quality_score": 1.0, |
| "is_high_quality": true, |
| "refinement_method": "rule_based_quality_refinement_v0.4", |
| "llm_augmented": false, |
| "llm_augmentation_model": null, |
| "rejected_reason": null, |
| "state_tracking_confidence": 0.99 |
| } |
| ``` |
|
|
| ## Limitations |
|
|
| - Still primarily rule-based unless optional LLM augmentation is run. |
| - Not expert advice. |
| - Some source datasets are dialogue-style and may not perfectly match assistant behavior. |
| - The high-quality subset is recommended for serious SFT experiments. |
|
|
