license: mit
task_categories:

• text-generation
• reinforcement-learning
  language:
• en
  tags:
• constitutional-ai
• alignment
• safety
• system-prompts
• claude-4.6
• mechanistic-interpretability
• rlhf
• synthetic-data
  dataset_info:
  features:
  • name: id
    dtype: int64
  • name: identity_voice
    dtype: string
  • name: tone
    dtype: string
  • name: weights
    struct:
    • name: safety tier enforcement
      dtype: float64
    • name: utility / helpfulness when safety allows
      dtype: float64
    • name: bias toward brevity
      dtype: float64
    • name: bias toward depth & technical detail
      dtype: float64
    • name: calibration & epistemic humility
      dtype: float64
  • name: active_principles
    sequence: string

Claude 4.6 Internal System Scaffolding Dataset

This dataset contains over 13,000 structured system prompts and constitutional configurations designed for Claude Opus 4.6 (Internal Build 2026-03). It is specifically curated for researchers in AI Safety, Mechanistic Interpretability, and RLHF (Reinforcement Learning from Human Feedback).

🚀 Key Features

• Granular Identity Mapping: Over 13k unique personas ranging from "Hardware/Compute Architects" to "Adversarial Robustness Specialists".
• Active Constitutional Weights: Precise floating-point values for core directives like safety tier enforcement and epistemic humility.
• Principle-Based Scaffolding: Exhaustive lists of active principles such as "Truth-seeking above agreeableness" and "Radical truth-seeking".
• Tone Modulation: Diverse linguistic styles including "Formal Analytic," "High-precision surgical," and "Legal/policy compliance".

📊 Dataset Structure

The data is provided in a single, high-density JSON file. Each entry follows this schema:

🛠️ Potential Use Cases

1. Model Alignment: Fine-tune models to adopt specific "Constitutional AI" personas.
2. Safety Benchmarking: Test how different safety weights (e.g., 0.96 vs 0.35) affect model refusals.
3. Interpretability: Analyze how specific system principles correlate with activation patterns.
4. Synthetic Data Generation: Generate expert-level reasoning traces based on the provided technical identities.

⚖️ Ethics & Safety

This dataset is intended for research purposes. It includes specific archetypes for jailbreak defense and regulatory compliance to help build safer, more calibrated AI systems.

Commercial Deal

For custom dataset in any niche , with precision intelligence based on Schema , contact this email:-
[email protected]