stindardlogic/email-writing-sft-100k

Email Writing SFT (100K)

100,000 ShareGPT conversations demonstrating professional email writing across 22 business contexts. Each example shows how to draft clear, purposeful emails that achieve their communication goal — from cold outreach to salary negotiations to apology emails.

Motivation

Email is the primary communication channel for most professional work, yet LLMs often produce emails that are:

• Too long: Including unnecessary preamble, excessive context, and trailing pleasantries that dilute the message
• Tone-blind: Using formal boilerplate for casual team messages, or overly casual language in sensitive professional situations
• Action-unclear: Burying the request or leaving the recipient uncertain about what's expected of them
• Context-generic: Failing to incorporate the specific details that make an email feel personal and relevant
• Stakes-unaware: Treating a salary negotiation the same as a meeting request

This dataset trains models to match tone, length, and directness to the specific professional context — and to include the specific details that make emails actually work.

Dataset Description

100,000 conversations across 22 email types:

Email Type Distribution

Email Type	Count	Description
client_communication	~12,000	Status updates, scope management, extensions
internal_request	~8,000	Salary negotiations, mentorship requests
cold_outreach	~4,000	Personalized pitch emails
follow_up	~4,000	Post-interview, post-demo follow-ups
negotiation	~4,000	Rate/contract negotiations
complaint	~4,000	Customer and vendor complaints
vendor_communication	~4,000	Contract terminations, vendor management
escalation	~4,000	Professional internal escalations
recommendation	~4,000	Employment recommendations
introduction	~4,000	Professional introductions
farewell	~4,000	Leaving a company gracefully
networking	~4,000	Conference follow-ups, staying in touch
congratulations	~4,000	Colleague promotions and wins
meeting_request	~4,000	Efficient meeting scheduling
hr_communication	~4,000	Workplace issue documentation
internal_announcement	~4,000	Policy changes, deadline updates
sales_followup	~4,000	Post-demo, post-meeting sales
customer_communication	~4,000	Apologies, service updates
thank_you	~4,000	Referrals, gratitude
professional_decline	~4,000	Job offer declines
policy_announcement	~4,000	Company policy changes

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Write a professional email to my team announcing that the project deadline has been moved up by two weeks..."
    },
    {
      "from": "gpt",
      "value": "Subject: Updated Project Deadline – Action Required

Hi team,

I wanted to give you a heads-up on a schedule change..."
    }
  ],
  "metadata": {
    "email_type": "internal_announcement",
    "context": "deadline_change",
    "tone": "professional_direct"
  },
  "id": "abc123"
}

Key Properties of Responses

1. Subject lines that get opened: Specific, action-oriented, and appropriately urgent. No generic "Following up" or "Quick question."

2. Front-loaded requests: The purpose of the email is clear in the first sentence or two. No one has to read the whole email to know what's being asked.

3. Appropriate length: Short emails for simple asks. Structured emails with headers for complex situations. Never longer than it needs to be.

4. Tone calibration: Warm with colleagues, assertive but respectful in negotiations, accountable in apologies, confident in cold outreach, gracious in declines.

5. Concrete specifics: Dates, numbers, names, and context that make the email feel real rather than generic. The salary negotiation email includes market research. The complaint email includes specific incident details.

6. Clean closes: Emails end when they've said what they needed to. No trailing summaries, no excessive sign-offs.

7. Difficult situations handled professionally: Scope creep addressed without passive aggression. Escalations framed constructively. Apologies that are accountable without excessive groveling.

Use Cases

• SFT fine-tuning for AI email assistants (Superhuman, Gmail, Outlook AI)
• Training enterprise writing assistants
• Building AI communication coaches for professionals
• Improving model performance on workplace communication tasks
• Training models for customer service and support platforms
• Fine-tuning models for sales enablement tools

License

Apache 2.0