stindardlogic/summarization-sft-100k

Summarization SFT (100K)

100,000 ShareGPT-format conversations covering document summarization across 9 source types and 5 summary styles. Trains models to summarize professional documents the way an expert human analyst would — identifying what matters, choosing the right format, and calibrating length to the task.

Motivation

Summarization is one of the most commercially deployed LLM capabilities, yet most summarization datasets train on news articles only. Real enterprise use cases involve email threads, meeting notes, incident reports, legal clauses, and Slack conversations — each with different structure, reading level, and output expectations. This dataset covers the full range.

Dataset Description

100,000 conversations across 9 source types and 5 summary styles:

Source Type Distribution

Source Type	Count	%
Email thread	~19,000	19%
Meeting notes	~14,300	14%
News article	~14,300	14%
Incident report	~9,500	10%
Academic abstract	~9,500	10%
Legal document	~9,500	10%
Product reviews	~9,500	10%
Technical report	~9,500	10%
Chat conversation	~4,800	5%

Summary Style Distribution

Style	Count	Description
bullets	~38,000	Structured bullet points with headings
executive_summary	~28,600	Paragraph form for leadership audiences
single_sentence	~14,300	One-sentence TL;DR
action_items	~9,500	Extracted action items with owners and dates
plain_language	~9,500	Technical content simplified for non-expert readers

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Summarize this email thread concisely:\n\n---\n**From:** Sarah Chen...\n**Subject:** Database migration status\n...\n"
    },
    {
      "from": "gpt",
      "value": "**Email Thread Summary**\n\nThe engineering and support teams are coordinating a PostgreSQL 11→15 database migration...\n\n**Open items**:\n1. Sarah to negotiate..."
    }
  ],
  "metadata": {
    "source_type": "email_thread",
    "summary_style": "executive_summary"
  },
  "id": "abc123"
}

Key Training Signals

1. Format selection: The response format matches the summary style requested — bullet points when asked, prose when asked, tables for structured data
2. Information priority: Key decisions, owners, and deadlines surface first; background context is compressed
3. Action item extraction: Responses identify responsible parties, tasks, and deadlines with high precision
4. Audience calibration: Technical content is simplified for plain_language style; preserved for bullets and executive_summary
5. Length discipline: single_sentence summaries are exactly one sentence; executive_summary runs 3–5 sentences

Source Type Details

• Email threads: Multi-party coordination, decisions, and follow-ups extracted from conversation flow
• Meeting notes: Unstructured notes reformatted into decisions, action items, and blockers
• Incident reports: Technical postmortems distilled for executive audiences and action tracking
• Academic abstracts: Research findings translated into accessible plain language
• Legal documents: Contract clauses explained without legalese
• Product reviews: Customer sentiment themes identified and structured for product teams
• Technical reports: Audit findings and recommendations structured by priority
• Chat conversations: Slack/Teams discussions condensed to decisions and next steps

Use Cases

• SFT fine-tuning for document summarization in enterprise applications
• Training models for email triage and meeting assistant products
• Legal tech applications requiring plain-language document summaries
• Customer success tools that summarize product review themes
• On-call tools that condense incident reports for engineering leadership

License

Apache 2.0