v0.2: Natural-language reasoning, cleaned public schema, SFT-ready

.gitattributes

@@ -61,3 +61,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 data/eval.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
+eval.jsonl filter=lfs diff=lfs merge=lfs -text
+high_quality_train.jsonl filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,78 +1,138 @@
----
-pretty_name: LifeTextSingleTurnStreamingCoT
-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
-- supervised-fine-tuning
-- life-assistant
----
 # LifeTextSingleTurnStreamingCoT
 
-LifeTextSingleTurnStreamingCoT is the canonical text/single-turn member of the Life Streaming CoT family. It supersedes the older `LifeStreamingCoT` name while preserving the existing v0.4.1 data and backward-compatible fields.
+**Version:** v0.2  
+**Modality:** text  
+**Turn type:** single_turn  
+**HF Repo:** [skyzhou06/LifeTextSingleTurnStreamingCoT](https://huggingface.co/datasets/skyzhou06/LifeTextSingleTurnStreamingCoT)
 
-## Summary
+## Dataset Overview
 
-- Modality: text
-- Turn type: single_turn
-- Version: v0.4.1
-- HF repo: `skyzhou06/LifeTextSingleTurnStreamingCoT`
-- Rows: 9322 total, 7457 train, 1865 eval
-- High-quality rows: 2570 train, 634 eval
-- Source distribution: `{"b-mc2/wikihow_lists": 622, "pietrolesci/multiwoz_all_versions": 2987, "pixelsandpointers/better_daily_dialog": 3713, "pixelsandpointers/empathetic_dialogues_for_lm": 2000}`
-- Category distribution: `{"daily_life": 2987, "information_extraction": 622, "social_communication": 5713}`
-- Length buckets: `{"short": 5575, "short_medium": 1345, "very_short": 2402}`
+LifeTextSingleTurnStreamingCoT is a single-turn text dataset for streaming chain-of-thought (CoT) training. Each example simulates gradually revealed information — the model sees the input in chunks and must produce streaming reasoning that evolves as more context becomes available. The dataset is designed for supervised fine-tuning (SFT) of language models that need to reason step-by-step over incrementally disclosed user requests.
+
+## Intended Use
+
+- Supervised fine-tuning of language models for streaming reasoning
+- Training models to produce natural-language chain-of-thought over partial inputs
+- Research on incremental understanding and progressive reasoning
+
+## Source Datasets and Licenses
+
+| Source | Domain | Rows | License |
+|--------|--------|------|---------|
+| b-mc2/wikihow_lists | how_to_guidance | 622 | varies |
+| pietrolesci/multiwoz_all_versions | task_oriented_assistant | 2,987 | varies |
+| pixelsandpointers/better_daily_dialog | daily_dialogue | 3,713 | varies |
+| pixelsandpointers/empathetic_dialogues_for_lm | emotional_support | 2,000 | varies |
+
+This dataset inherits licenses from its upstream sources. Users are responsible for complying with source dataset terms.
 
 ## Schema
 
-Core legacy fields include `id`, `domain`, `source_dataset`, `instruction`, `context`, `context_chunks`, `streaming_reasoning`, `deep_reasoning`, `answer`, `messages`, `text`, `quality_flags`, `quality_score`, `is_high_quality`, and `split`.
+Each row contains:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `id` | string | Unique row identifier |
+| `version` | string | Dataset version (`v0.2`) |
+| `split` | string | `train` or `eval` |
+| `modality` | string | `text` |
+| `turn_type` | string | `single_turn` |
+| `input` | object | Input fields (instruction, context, token count) |
+| `output` | object | Target fields (deep_reasoning, answer) |
+| `streaming` | object | Streaming reasoning and checkpoints |
+| `taxonomy` | object | Category, subcategory, intent type |
+| `quality` | object | Quality flags, score, high-quality marker |
+| `metadata` | object | Source dataset, original version |
+
+### Input Fields
+
+- `input.instruction`: Task instruction
+- `input.context`: Full input context
+- `input.input_token_count`: Approximate token count
+- `input.length_bucket`: Length category
+
+### Output Fields (Recommended SFT Targets)
+
+- `output.deep_reasoning`: **Natural-language deep reasoning** — summarizes the user goal, known constraints, missing information, and answer plan.
+- `output.answer`: The final assistant response.
+
+### Streaming Fields
+
+- `streaming.streaming_reasoning`: **Natural-language streaming reasoning** — describes how understanding evolves as each input chunk is revealed.
+- `streaming.checkpoints`: Per-chunk checkpoints with individual reasoning.
+- `streaming.trace`: Original slot-style trace (auxiliary, not recommended for training).
 
-Normalized backward-compatible fields were added where feasible: `modality`, `turn_type`, `taxonomy`, `input`, `streaming`, `output`, and `quality`.
+### Legacy / Trace Fields
 
-## Taxonomy
+The original dataset (v0.4.1) used slot-style reasoning (e.g., `domain=restaurant; goal=search`). In v0.2, these have been converted to natural language. The original traces are preserved under:
 
-Rows include normalized `taxonomy.category`, `taxonomy.subcategory`, `taxonomy.intent_type`, and `taxonomy.difficulty`. Existing `domain` remains available for backward compatibility.
+- `streaming.trace`: Original slot-style streaming reasoning
+- `output.deep_reasoning_trace`: Original slot-style deep reasoning
 
-## Streaming and Deep Reasoning
+These trace fields are **not recommended** as SFT targets.
 
-Streaming reasoning is deterministic, selective, and chunk-aligned. Deep reasoning is a compact full-context summary. No `sft_messages` field is required.
+## Recommended Fields for SFT
+
+For supervised fine-tuning:
+
+- **Input:** `input.context`
+- **Streaming reasoning:** `streaming.streaming_reasoning` or `streaming.checkpoints[*].streaming_reasoning`
+- **Deep reasoning:** `output.deep_reasoning`
+- **Answer:** `output.answer`
+
+## Natural-Language Reasoning (v0.2)
+
+v0.2 uses **natural-language reasoning** as the primary training target. All slot-style reasoning has been converted to flowing natural language.
 
 ## Quality Filters
 
-The release keeps quality scores, high-quality split files, quality flags, source attribution, and validation scripts. The high-quality config is recommended for stricter SFT experiments.
+- `quality.is_high_quality`: Boolean, stricter quality checks
+- `quality.quality_score`: Numeric score (0.0–1.0)
+- `quality.quality_flags`: List of quality flags
+- High-quality subsets: `high_quality_train.parquet`, `high_quality_eval.parquet`
+
+## Split Counts
+
+| Split | Rows |
+|-------|------|
+| train | 7,457 |
+| eval | 1,865 |
+| high_quality_train | 2,570 |
+| high_quality_eval | 634 |
+| **Total** | **9,322** |
 
-## How to use for SFT
+## Category Distribution
 
-- Input: `input.instruction` plus `input.context` or legacy `instruction` plus `context`.
-- Target: `output.answer` or legacy `answer`.
-- Optional reasoning target: `streaming.streaming_reasoning`, `output.deep_reasoning`, then `output.answer`.
+| Category | Rows |
+|----------|------|
+| social_communication | 5,713 |
+| daily_life | 2,987 |
+| information_extraction | 622 |
+
+## Known Limitations
+
+- Primarily rule-based. Reasoning quality varies with source data.
+- The high-quality subset is recommended for serious SFT experiments.
+- Some source datasets are dialogue-style and may not match ideal assistant behavior.
+- Not intended for expert medical, legal, financial, or safety-critical advice.
+
+## How to Load
 
 ```python
 from datasets import load_dataset
 
-full = load_dataset("skyzhou06/LifeTextSingleTurnStreamingCoT", "default")
-hq = load_dataset("skyzhou06/LifeTextSingleTurnStreamingCoT", "high_quality")
+dataset = load_dataset("skyzhou06/LifeTextSingleTurnStreamingCoT", split="train")
+print(dataset[0])
 ```
 
-## Limitations
+## How to Map to SFT
 
-- Reasoning is deterministic/rule-based unless optional LLM augmentation is run separately.
-- Some rows originate from dialogue-style sources and may not perfectly match direct assistant behavior.
-- Not intended for expert medical, legal, financial, emergency, or safety-critical advice.
+```python
+def format_sft_sample(row):
+    return {
+        "messages": [
+            {"role": "user", "content": row["input"]["context"]},
+            {"role": "assistant", "content": row["output"]["deep_reasoning"] + "\n\n" + row["output"]["answer"]}
+        ]
+    }
+```

build_report.json

@@ -0,0 +1,119 @@
+{
+  "dataset": "LifeTextSingleTurnStreamingCoT",
+  "version": "v0.2",
+  "modality": "text",
+  "turn_type": "single_turn",
+  "build_timestamp": "2026-06-06T22:01:31.021319",
+  "row_counts": {
+    "train": 7457,
+    "eval": 1865,
+    "high_quality_train": 2570,
+    "high_quality_eval": 634,
+    "high_quality_total": 3204,
+    "total": 9322
+  },
+  "reasoning_conversion": {
+    "slot_style_sr_before": 200,
+    "slot_style_dr_before": 62,
+    "slot_style_sr_after": 0,
+    "slot_style_dr_after": 0,
+    "natural_language_pass": true
+  },
+  "schema_changes": {
+    "fields_removed": [
+      "reasoning_token_budget",
+      "chunk_labels",
+      "skip_chunks",
+      "skip_reasons",
+      "chunk_split_count",
+      "original_num_chunks",
+      "generation_method",
+      "reasoning_policy",
+      "chunking_method",
+      "refinement_method",
+      "llm_augmented",
+      "llm_augmentation_model",
+      "state_tracking_confidence",
+      "rejected_reason"
+    ],
+    "fields_before_count": 40,
+    "fields_after_count": 23,
+    "fields_before": [
+      "answer",
+      "chunk_labels",
+      "chunk_split_count",
+      "chunking_method",
+      "context",
+      "context_chunks",
+      "deep_reasoning",
+      "domain",
+      "generation_method",
+      "id",
+      "input",
+      "instruction",
+      "is_high_quality",
+      "language",
+      "llm_augmentation_model",
+      "llm_augmented",
+      "messages",
+      "modality",
+      "num_chunks",
+      "original_num_chunks",
+      "output",
+      "quality",
+      "quality_flags",
+      "quality_score",
+      "reasoning_policy",
+      "reasoning_token_budget",
+      "refinement_method",
+      "rejected_reason",
+      "response",
+      "skip_chunks",
+      "skip_reasons",
+      "source_dataset",
+      "split",
+      "state_tracking_confidence",
+      "streaming",
+      "streaming_reasoning",
+      "taxonomy",
+      "text",
+      "turn_type",
+      "version"
+    ],
+    "fields_after": [
+      "answer",
+      "context",
+      "deep_reasoning",
+      "domain",
+      "id",
+      "input",
+      "instruction",
+      "is_high_quality",
+      "language",
+      "messages",
+      "metadata",
+      "modality",
+      "output",
+      "quality",
+      "quality_flags",
+      "quality_score",
+      "source_dataset",
+      "split",
+      "streaming",
+      "streaming_reasoning",
+      "taxonomy",
+      "turn_type",
+      "version"
+    ]
+  },
+  "source_distribution": {
+    "pixelsandpointers/better_daily_dialog": 3713,
+    "pietrolesci/multiwoz_all_versions": 2987,
+    "pixelsandpointers/empathetic_dialogues_for_lm": 2000,
+    "b-mc2/wikihow_lists": 622
+  },
+  "hq_before_vs_after": {
+    "before": 3204,
+    "after": 3204
+  }
+}

eval.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57c3b1696abfce5146a6ac04b06b97c23c4459460e8ca2738f49c8168ae30ecb
+size 11032440

eval.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3945706b6817f157201baa0fab37e99921f88183f16f60c941565f17087ebb0d
+size 2750452

high_quality_eval.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7898581422ea0eb41bff56e642a83b4ffa549a9c2a3d10ea804bfdb9faf5ac2c
+size 908174

high_quality_train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f068d8de07fb01436c0b9528548f97615d18e01b81a515af1a1ae7f6669ebf15
+size 16550697

high_quality_train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17012e97fe1b0e352f7ed5e3fc69f646f2aa3b91c3cdfd8cc4257035ffcb51f9
+size 3537033

inspect_report.json

@@ -0,0 +1,99 @@
+{
+  "dataset": "LifeTextSingleTurnStreamingCoT",
+  "version": "v0.2",
+  "modality": "text",
+  "turn_type": "single_turn",
+  "inspect_timestamp": "2026-06-06T22:09:59.314820",
+  "checks": {
+    "data_load": {
+      "status": "PASS",
+      "train_rows": 7457,
+      "eval_rows": 1865,
+      "total_rows": 9322
+    },
+    "required_fields": {
+      "status": "PASS",
+      "missing": [],
+      "total_fields": 23,
+      "all_fields": [
+        "answer",
+        "context",
+        "deep_reasoning",
+        "domain",
+        "id",
+        "input",
+        "instruction",
+        "is_high_quality",
+        "language",
+        "messages",
+        "metadata",
+        "modality",
+        "output",
+        "quality",
+        "quality_flags",
+        "quality_score",
+        "source_dataset",
+        "split",
+        "streaming",
+        "streaming_reasoning",
+        "taxonomy",
+        "turn_type",
+        "version"
+      ]
+    },
+    "forbidden_fields": {
+      "status": "PASS",
+      "forbidden_found": []
+    },
+    "answer_nonempty": {
+      "status": "PASS",
+      "train_nonempty": 7457,
+      "train_empty": 0,
+      "eval_nonempty": 1865,
+      "eval_empty": 0
+    },
+    "natural_language_reasoning": {
+      "status": "PASS",
+      "streaming_slot_style_count": 0,
+      "deep_slot_style_count": 0,
+      "sample_size": 500,
+      "streaming_pass": true,
+      "deep_pass": true
+    },
+    "sft_readiness": {
+      "status": "PASS",
+      "has_input_field": true,
+      "has_output_field": true,
+      "has_deep_reasoning_in_output": true
+    },
+    "quality_fields": {
+      "status": "PASS",
+      "has_is_high_quality": true,
+      "has_quality_score": true
+    },
+    "source_license": {
+      "status": "PASS",
+      "has_source": true,
+      "has_license": true
+    },
+    "taxonomy": {
+      "status": "PASS",
+      "has_taxonomy_field": true,
+      "category_nonempty_pct": 100
+    },
+    "splits": {
+      "train_split_distribution": {
+        "train": 7457
+      },
+      "eval_split_distribution": {
+        "eval": 1865
+      }
+    },
+    "high_quality": {
+      "has_high_quality_split": true,
+      "high_quality_rows": 3204
+    }
+  },
+  "publishability_status": "PASS",
+  "publishability_limitations": []
+}

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a10e6a5fa2c97d39c6713ba14e9507ab461aefc1f440ce63bd260d339b8682c
+size 44469780

train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:776cbf3a3ba79ae1d6627de0f5f83be06f60274ab6e4fd20ac7fe8feb5fded13
+size 10983685