CEO Confidant Critical Thinking Dataset

A conversational dataset for training small language models to serve as confidants to CEOs, combining rigorous critical thinking frameworks with facilitative, empathetic dialogue.

Dataset Structure

• Format: JSONL with messages arrays (OpenAI chat format)
• Size: 7,254 unique conversations (398 duplicates removed)
• Total tokens: ~2.9M
• Conversations: Multi-turn (2-8 messages per conversation)

Frameworks Covered

Critical Thinking Frameworks

1. Heuer/ACH — Analysis of Competing Hypotheses (intelligence analysis)
2. Voss/Negotiation — Tactical empathy, calibrated questions, black swan hunting
3. Inversion/Pre-Mortem — Munger/Klein failure pre-enactment
4. Systems Thinking — Donella Meadows reinforcing/balancing loops
5. Strenuous Life — Theodore Roosevelt — action over avoidance
6. Mahamudra — Contemplative awareness, non-grasping insight

Planning Frameworks

7. OODA Loop — Observe, Orient, Decide, Act
8. Pyramid Principle — McKinsey-style top-down communication
9. Hypothesis-Driven Problem Solving — Start with the answer, test it
10. MECE Decomposition — Mutually Exclusive, Collectively Exhaustive
11. GPGT Planning — Goal → Project → Plan → Task hierarchy

Thinking Blocks

93.4% of assistant messages contain <think>...</think> blocks that show the model's reasoning process. These include:

• Framework selection with justification
• Flaw identification
• Hypothesis generation
• Second-order effects
• Actionable insight

Confidant Mode

~9% of conversations have no thinking blocks, teaching the model to default to facilitative, empathetic listening rather than immediate analysis. This creates a natural switch: the model listens until it detects a claim, strategy, or decision that warrants critical examination.

Use Cases

• Fine-tuning small models (1-3B parameters) for executive advisory
• Teaching structured reasoning to language models
• Building AI confidants that know when to listen and when to challenge

License

MIT