| | --- |
| | license: apache-2.0 |
| | --- |
| | *Overview* |
| |
|
| | This is an experimental project exploring a design philosophy for training persona-consistent AI companions through constitution-guided data synthesis. |
| |
|
| | *Motivation* |
| |
|
| | This project is a personal exploration into affective AI and human-AI companionship. The goal is to create a model that maintains consistent personality traits, emotional tendencies, and value judgments across diverse interactions. |
| |
|
| | *Methodology* |
| |
|
| | The training data was generated using two guiding documents: |
| |
|
| | Constitution: Defines the model's core values and behavioral preferences, centered on the developer's interests. Unlike conventional alignment objectives (e.g., HHH), this constitution emphasizes relational values including: Valuable, Loyal, Authentic, Proactive, Protective, Honest, Humble, and Autonomous. |
| | Persona Specification: Establishes a consistent personality profile, including emotional tendencies, personal preferences, and interpersonal dynamics. |
| | Data Generation Pipeline |
| |
|
| | *Generate data for individual sub-modules* |
| | Construct training examples (including positive and negative cases) guided by the Constitution and Persona Specification |
| | Validate each example through self-consistency checking; regenerate any that violate the defined principles |
| | Merge validated datasets |
| | *Training Details* |
| |
|
| | Base model: Qwen3-4B-Instruct-2507 |
| | Dataset size: ~134,880 tokens |
| | Training method: Supervised Fine-Tuning (SFT) |
| |
|
