| --- |
| language: |
| - en |
| license: apache-2.0 |
| size_categories: |
| - 100K<n<1M |
| task_categories: |
| - text-generation |
| pretty_name: Data Extraction SFT (100K) |
| tags: |
| - information-extraction |
| - structured-extraction |
| - nlp |
| - enterprise |
| - document-parsing |
| - json-extraction |
| - table-extraction |
| - ner |
| - sft |
| - supervised-fine-tuning |
| - synthetic |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data-extraction-sft-100k.jsonl |
| --- |
| |
| # Data Extraction SFT (100K) |
|
|
| 100,000 ShareGPT conversations demonstrating structured information extraction from unstructured text. Each example takes a real-world document (invoice, contract, resume, research abstract, meeting notes, log files) and extracts the relevant information into JSON, markdown tables, or other structured formats. |
|
|
| ## Motivation |
|
|
| Information extraction is one of the highest-value NLP tasks in enterprise settings. Common model failures include: |
|
|
| - **Format drift**: Asked for JSON, produces a mix of prose and JSON |
| - **Missed fields**: Extracts some fields but silently omits others |
| - **Hallucinated data**: Infers or invents values not present in the source |
| - **Wrong output format**: Produces a table when JSON was requested, or vice versa |
| - **Poor handling of ambiguity**: Doesn't flag when source data is incomplete or contradictory |
| - **Inconsistent structure**: JSON with inconsistent key naming, tables with misaligned columns |
|
|
| This dataset trains models to extract accurately, completely, and in the exact format requested. |
|
|
| ## Dataset Description |
|
|
| **100,000 conversations** across 23 extraction types: |
|
|
| ### Extraction Type Distribution |
|
|
| | Extraction Type | Source Format | Output Format | |
| |---|---|---| |
| | `contact_information` | Email signature | JSON | |
| | `financial_data` | Earnings report | Table | |
| | `job_posting` | Job description | JSON | |
| | `dates_events` | Project timeline | Table | |
| | `contract_terms` | Legal clause | Bullet list | |
| | `product_specifications` | Product description | Table | |
| | `action_items` | Meeting notes | Table | |
| | `named_entities` | News article | Categorized list | |
| | `invoice_parsing` | Invoice | JSON | |
| | `prescription_data` | Prescription | JSON | |
| | `network_entities` | Server log | Categorized list | |
| | `requirements_classification` | Requirements doc | Classified list | |
| | `clinical_trial_data` | Research abstract | JSON | |
| | `sentiment_analysis` | Customer review | Analysis | |
| | `recipe_ingredients` | Recipe | JSON | |
| | `meeting_decision` | Meeting transcript | Structured analysis | |
| | `api_error_codes` | API documentation | Table | |
| | `business_metrics` | KPI dashboard | JSON | |
| | `resume_parsing` | Resume | JSON | |
| | `compliance_violations` | Audit report | Table | |
| | `real_estate_listing` | Property listing | JSON | |
| | `performance_review` | HR document | Structured summary | |
| | `mathematical_formulas` | Scientific text | Structured analysis | |
|
|
| ## Format |
|
|
| ```json |
| { |
| "conversations": [ |
| { |
| "from": "human", |
| "value": "Extract all contact information from the following email signature and return it as JSON: |
| |
| --- |
| Dr. Sarah Mitchell... |
| ---" |
| }, |
| { |
| "from": "gpt", |
| "value": "```json |
| {"name": "Dr. Sarah Mitchell", "title": "Chief Research Officer"...} |
| ```" |
| } |
| ], |
| "metadata": { |
| "extraction_type": "contact_information", |
| "source_format": "email_signature", |
| "output_format": "json" |
| }, |
| "id": "abc123" |
| } |
| ``` |
|
|
| ## Key Properties of Responses |
|
|
| **1. Format fidelity**: JSON when JSON is requested, markdown tables when tables are requested. Code blocks used correctly. Keys are consistently named. |
|
|
| **2. Complete extraction**: All entities in the source are extracted. Nothing is silently skipped. When the source is ambiguous, the extraction notes the ambiguity. |
|
|
| **3. No hallucination**: Only information present in the source is extracted. Missing values are represented as `null` or omitted, not invented. |
|
|
| **4. Appropriate structure**: Nested JSON for hierarchical data, flat tables for tabular data, categorized lists for multi-type entity extraction. |
|
|
| **5. Relevant metadata**: Extracted data includes context that makes it actionable (e.g., units for measurements, calculation details for formulas, severity indicators for compliance violations). |
|
|
| **6. Real-world document types**: Source documents represent actual professional formats — invoices with realistic line items, resumes with realistic career progressions, API docs with standard error code patterns. |
|
|
| ## Use Cases |
|
|
| - SFT fine-tuning for document processing pipelines |
| - Training AI for enterprise automation (invoice processing, contract review, HR document parsing) |
| - Building RAG pre-processing models that extract structured facts from documents |
| - Improving model performance on information extraction benchmarks |
| - Training models for legal tech, HR tech, and fintech applications |
| - Building AI assistants for analysts who work with unstructured data |
|
|
| ## License |
|
|
| Apache 2.0 |
|
|
