Add dataset card

README.md

@@ -0,0 +1,110 @@
+---
+language:
+- en
+license: apache-2.0
+size_categories:
+- 100K<n<1M
+task_categories:
+- text-generation
+pretty_name: Document Summarization DPO (100K)
+tags:
+- document-summarization
+- summarization
+- dpo
+- direct-preference-optimization
+- synthetic
+- enterprise
+- information-extraction
+- business-documents
+- rlhf
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: document-summarization-dpo-100k.jsonl
+---
+
+# Document Summarization DPO (100K)
+
+100,000 DPO (Direct Preference Optimization) preference pairs for training models to summarize business and professional documents with precision, structure, and analytical depth.
+
+## Motivation
+
+Document summarization is one of the highest-value enterprise AI applications — analysts, lawyers, product managers, and executives use AI to process reports, contracts, and research daily. Models commonly fail by:
+
+- **Losing quantitative data**: Summarizing "the company performed well" instead of "revenue grew 23% YoY to $4.2B, with operating margin expanding 180bps to 18.3%"
+- **Vague prose instead of structure**: Running text instead of tables, bullets, or sections that match the source document's information hierarchy
+- **Adding no analytical value**: Copying sentences instead of synthesizing, deriving implications, or flagging key risks
+- **Hallucinating or omitting key clauses**: Legal and compliance documents require verbatim fidelity to numbers, dates, and obligations
+- **Ignoring document type conventions**: An earnings report summary should differ from a meeting minutes summary in structure, depth, and emphasis
+
+This dataset trains models to summarize with the discipline of a senior analyst: preserving all quantitative data, using structured output, deriving explicit insights, and flagging risks.
+
+## Dataset Description
+
+**100,000 DPO preference pairs** across 16 document types:
+
+### Document Types
+
+| Type | Description |
+|---|---|
+| `earnings_report` | Quarterly/annual financial performance |
+| `research_paper_abstract` | Scientific and academic research |
+| `legal_contract` | Contracts, agreements, terms |
+| `incident_report` | Technical outages and post-mortems |
+| `meeting_minutes` | Decisions and action items from meetings |
+| `policy_document` | Internal policies and procedures |
+| `product_spec` | Technical specifications and requirements |
+| `scientific_article` | Peer-reviewed research papers |
+| `customer_complaint` | Support escalations and complaints |
+| `technical_documentation` | API docs, system documentation |
+| `strategic_plan` | Business strategy and OKRs |
+| `job_description` | Role requirements and responsibilities |
+| `audit_report` | Internal/external audit findings |
+| `news_article` | Business and industry news |
+| `performance_review` | Employee evaluations and feedback |
+| `medical_report` | Clinical notes and medical assessments |
+
+## Format
+
+```json
+{
+  "prompt": "Summarize the following earnings report:\n\n[document text]",
+  "chosen": "**Q2 2026 Financial Summary**\n\n**Revenue:** $4.2B (+23% YoY, +4.1% QoQ)...\n\n**Key Insights:** Operating leverage is expanding faster than consensus expected...",
+  "rejected": "The company reported strong financial results for Q2 2026. Revenue and profits both increased compared to last year. The company expressed optimism about future performance.",
+  "metadata": {
+    "doc_type": "earnings_report",
+    "chosen_rationale": "Preserves all quantitative data, uses structured format, adds derivable analytical insights",
+    "rejected_rationale": "Vague prose loses all specific numbers and adds no analytical value"
+  },
+  "id": "abc123"
+}
+```
+
+## Key Properties
+
+**Chosen responses:**
+- **All numbers preserved**: Every percentage, dollar figure, date, and metric from the source appears in the summary
+- **Structured output**: Uses tables, bullets, and section headers appropriate to the document type
+- **Analytical synthesis**: Derives implications, flags risks, and highlights what changed vs. expectation — not just what happened
+- **Appropriate length**: Neither too terse (losing data) nor too verbose (adding noise)
+- **Document-type conventions**: Earnings summaries include segment breakdowns; legal summaries flag obligations and risks; meeting minutes lead with decisions and owners
+
+**Rejected responses:**
+- Vague prose that substitutes qualitative language for quantitative data
+- Hallucinated or omitted specifics
+- Flat summaries that add no synthesis beyond restating what the document said
+- Unstructured text for documents that have clear hierarchical information
+
+## Use Cases
+
+- DPO/RLHF fine-tuning for enterprise document AI (legal tech, finance, healthcare)
+- Training summarization models for analyst and knowledge worker tools
+- Building AI assistants for contract review, earnings analysis, and report processing
+- Improving model performance on structured information extraction
+- Fine-tuning models for document intelligence platforms (M&A, compliance, audit)
+- Training medical AI for clinical note summarization
+
+## License
+
+Apache 2.0