stindardlogic/communication-adaptation-sft-100k

Communication Adaptation SFT (100K)

100,000 ShareGPT conversations demonstrating skilled communication style adaptation across 15 task types. Each example shows how to take the same underlying content and adjust register, technical depth, length, and framing for different audiences and purposes.

Motivation

Communication adaptation is a core professional skill that LLMs often handle clumsily. Common failures:

• Technical monologue: explaining cloud storage to a grandmother the same way as to a developer
• Tone blindness: maintaining formal register in a casual Slack message, or vice versa
• Length calibration errors: giving a one-page answer when a sentence was needed, or a bullet when a paragraph was needed
• Audience blindness: pitching the same product the same way to an engineer as to a CFO
• Missed framing opportunities: presenting a budget cut identically to people who need to hear it as an opportunity vs. a risk

This dataset trains models to identify the audience, purpose, and context of a communication task -- and adjust accordingly.

Dataset Description

100,000 conversations across 15 adaptation types:

Adaptation Type Distribution

Adaptation Type	Count	%	Description
technical_to_nontechnical	~30,000	30%	Explain technical topics to non-expert audiences
reading_level	~5,000	5%	Same topic at multiple expertise levels
executive_summary	~5,000	5%	Condense technical reports for C-suite
formal_to_casual	~5,000	5%	Convert formal text to conversational
casual_to_formal	~5,000	5%	Convert casual text to professional
expert_to_beginner	~5,000	5%	Simplify expert knowledge for newcomers
condensed_to_expanded	~5,000	5%	Expand bullet points into full prose
expanded_to_condensed	~5,000	5%	Distill long text into TL;DR
persuasive_vs_informative	~5,000	5%	Same content, different rhetorical purpose
cultural_register	~5,000	5%	Adapt one event for internal/external/public audiences
audience_expertise	~5,000	5%	Same topic at analyst vs junior dev level
length_calibration	~5,000	5%	One-sentence, paragraph, and full-page versions
register_formality	~5,000	5%	Peer vs. junior vs. formal review tone
instructional_scaffolding	~5,000	5%	Instructions at novice, intermediate, and expert levels
persuasive_framing	~5,000	5%	Same decision framed as challenge, opportunity, or risk

Subject Domains

Business communications, technical explanations (APIs, databases, encryption, machine learning, version control, cloud storage), financial concepts, educational content, incident management, product documentation, HR and management communication.

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Explain API rate limiting to a non-technical executive who needs to understand why our integration is sometimes slow."
    },
    {
      "from": "gpt",
      "value": "Think of our API connection like a water tap connected to a shared pipe system. The vendor limits how much water any one customer can draw per minute..."
    }
  ],
  "metadata": {
    "adaptation_type": "technical_to_nontechnical",
    "topic": "explain_api_rate_limiting",
    "audience": "non_technical_executive",
    "original_context": "technical"
  },
  "id": "abc123"
}

Key Properties of Responses

1. Audience vocabulary calibration: Technical terms are replaced with analogies for non-expert audiences, retained and built upon for expert audiences. The same underlying concept uses different vocabulary at different levels.

2. Appropriate analogy use: Non-technical explanations use domain-appropriate analogies (Photoshop History for developers explaining Git to a designer; filing cabinet for explaining database indexes to a product manager). Analogies break down gracefully rather than overextending.

3. Length matching: Responses are calibrated to the scope requested. A TL;DR is actually 2-3 sentences. A one-page explanation fills a page. Instructions for a novice include every step; instructions for an expert omit what they already know.

4. Register consistency: Once a register is established (casual, formal, literary), it's maintained throughout the response. Mixed registers are avoided.

5. Framing without distortion: Persuasive reframings (budget cut as opportunity) don't invent facts -- they recontextualize the same information differently. The underlying truth is preserved.

6. Multi-version responses: When asked for the same content at multiple levels, each version is genuinely different -- not just more or fewer words, but different vocabulary, examples, structure, and assumptions.

7. Scope awareness: Executive summaries omit technical detail; technical documentation is precise. Each format serves its purpose.

Use Cases

• SFT fine-tuning for AI writing assistants (enterprise, consumer)
• Training AI communication coaches
• Building multi-register language models for professional tools
• Training models for business communication platforms (email, Slack, Notion AI)
• Improving model performance on audience-adaptation tasks
• Building AI that can serve both technical and non-technical users in the same product

License

Apache 2.0