stindardlogic/meeting-summarization-sft-100k

Meeting Summarization SFT (100K)

100,000 ShareGPT conversations demonstrating structured meeting summarization across 22 meeting types. Each example converts a realistic meeting transcript into a well-organized summary with key decisions, action items, and discussion notes — in the format that professional teams actually use.

Motivation

Meeting transcription tools (Otter.ai, Fireflies, Zoom AI) generate raw text but struggle to produce usable summaries. Common model failures:

• Chronological dumps: Restating what was said in order instead of synthesizing decisions and actions
• Missing ownership: Noting that something should be done without attributing it to a specific person
• Wrong format: Producing a paragraph summary when the meeting had 8 distinct action items that need a table
• No signal prioritization: Burying the key decision in the middle of a summary with the same weight as small talk
• Format inconsistency: Different output structures for different meetings instead of adapting to the meeting type

This dataset trains models to synthesize rather than transcribe, extract rather than summarize, and format appropriately for the type of meeting.

Dataset Description

100,000 conversations across 22 meeting types:

Meeting Type	Domain	Output Format
product_planning	Product	Decisions + action table + discussion notes
executive_review	Strategy	Executive summary with metrics table
standup	Engineering	Status table + blocker list
design_review	Product/Design	Decision log + required revisions + open items
post_mortem	Engineering	Root cause + timeline + action items
sales_review	Sales	Pipeline table + deal notes + actions
hiring_debrief	HR	Candidate comparison + decision + offer details
okr_review	Strategy	Scorecard table + carryovers
customer_success	CS	Issue table + expansion opportunity + actions
roadmap_planning	Product	Initiative table + resource allocation + rationale
board_meeting	Executive	Metrics summary + strategic updates + board notes
all_hands	Company	Announcement + changes + Q&A highlights
architecture_review	Engineering	Proposal + discussion points + revised sequence
one_on_one	Management	Status + feedback + open items
budget_review	Finance	Budget comparison table + department notes
legal_compliance	Legal	Compliance status + gap analysis + actions
sprint_retrospective	Agile	What went well + issues + process changes
partnership_discussion	BD	Integration concept + commercial terms + next steps
research_readout	UX/Research	Findings + recommendations + feasibility
vendor_evaluation	Procurement	Comparison table + decision + rationale
risk_review	Risk	Risk register + remediation tracking
investor_update	IR	Metrics + growth + strategic updates + asks

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Summarize the following meeting transcript into a structured summary...\n\n[TRANSCRIPT]\n..."
    },
    {
      "from": "gpt",
      "value": "## Meeting Summary\n\n**Date:** [Meeting Date]\n**Attendees:** ...\n\n### Key Decisions\n..."
    }
  ],
  "metadata": {
    "meeting_type": "product_planning",
    "output_type": "structured_summary",
    "domain": "product_development"
  },
  "id": "abc123"
}

Key Properties of Responses

1. Decisions, not discussion: Summaries lead with what was decided, not what was discussed. A product planning meeting output leads with "iOS-first launch approved" not "the team talked about launch timing."

2. Action items with owners and deadlines: Every action item has a clear owner (person, not team) and a deadline. Vague "team to follow up" items don't appear.

3. Format matched to meeting type: Standups get status tables. Board meetings get executive summaries. Post-mortems get root cause + timeline + remediation. The format serves the reader, not the transcription.

4. Realistic transcripts: Source transcripts include interruptions, side discussions, filler conversation, and incomplete sentences — the kind of input real transcription tools produce.

5. Numerical precision: Financial figures, percentages, dates, and counts from the transcript are preserved exactly in the summary.

6. Open items vs. closed items: Decisions made are clearly separated from topics that need follow-up. Readers know what's settled and what isn't.

Use Cases

• SFT fine-tuning for meeting AI tools (Notion AI, Microsoft Copilot, Otter.ai, Fireflies)
• Training enterprise AI assistants for business communication
• Building AI note-takers for sales, product, and engineering teams
• Improving model performance on summarization benchmarks
• Training models for enterprise productivity platforms
• Building AI for board and investor communication workflows

License

Apache 2.0