Chordia: High-Precision AI Emotional Dynamics Core

Plucking the strings of the mind, analyzing the instantaneous sense of resonance.

A deep learning-based AI emotional evolution prediction system. This project utilizes a Multi-Layer Perceptron (MLP) to fit emotional state transitions during interactions, providing AI characters with sub-millisecond physiological and emotional response capabilities.

🎯 Core Architecture: Decoupling Perception and Logic

This project adopts a dual-architecture of "Core Perception Prediction + Dynamic Logic Mapping":

• Perception Kernel (MLP): Focuses on predicting the trend of core emotional polarity (PAD) transitions.
• Runtime Mapping (Engine): Derives pressure values through linear scaling and physical formulas, achieving dynamic personality adjustment.

📦 Version Information

Current Version: v0.0.1-alpha (Chordia-P100)

This version consists of the optimal weights extracted from our training machine, fully verified and tested for reproducibility, offering the best stability and prediction accuracy.

Training Environment

The model was trained in the following hardware environment:

Component	Specification
GPU	NVIDIA Tesla P100-PCIE-16GB (16GB HBM2)
CUDA Version	12.8
Driver Version	570.169
Compute Capability	6.0 (Pascal Architecture)

Reproducibility Guarantee

• ✅ Code Reproducibility: 100% - All training code is open-sourced.
• ✅ Configuration Reproducibility: 100% - Training configuration files are identical.
• ✅ Weight Consistency: Identical to the version on the training machine.
• ✅ Performance Verification: Achieves the same metrics on the standard test set.
• 📄 Training Logs:
  • chordia_v0.0.1-alpha_training.log - Training summary (1.7KB)
  • chordia_v0.0.1-alpha_training_full.log - Full training record (604KB)

🚀 Key Performance Indicators (Benchmark)

After 500-600 epochs of training, the model demonstrates strong fitting capabilities:

Dimension	$R^2$ (Explained Variance)	MAE (Mean Absolute Error)	Psychological Significance
ΔP (Pleasure)	0.488	0.123	Empathy: Accurately perceives likes and dislikes from environmental stimuli.
ΔA (Arousal)	0.550	0.112	Expressiveness: Precisely predicts emotional tension and reaction intensity.
ΔD (Dominance)	0.058	0.097	Consistency: Maintains personality background, ensuring dominance stability.

💡 Design Philosophy: The low $R^2$ for $\Delta D$ is intended to ensure the long-term stability of the AI's dominance, avoiding unnatural fluctuations in personality traits due to random inputs.

Metric	Value	Description
Test MAE	0.111	Overall prediction error
Test $R^2$ (Mean)	0.366	Average explained variance
Test $R^2$ (Robust)	0.447	Robust explained variance
Validation Loss	0.023	Best validation set loss
Inference Latency	< 1ms	Single inference time

• Training Stability: Uses AdamW optimizer (lr=0.0005) combined with Cosine Annealing learning rate scheduling (T_max=600), and an early stopping mechanism (patience=150) to prevent overfitting.

📊 Input/Output Specifications

Input Features (7 Dimensions)

Feature Name	Description	Range
user_pleasure	User Pleasure	[-1.0, 1.0]
user_arousal	User Arousal	[-1.0, 1.0]
user_dominance	User Dominance	[-1.0, 1.0]
vitality	AI Character Physiological Vitality	[0.0, 100.0]
current_pleasure	AI Current Pleasure	[-1.0, 1.0]
current_arousal	AI Current Arousal	[-1.0, 1.0]
current_dominance	AI Current Dominance	[-1.0, 1.0]

Output Predictions (3 Dimensions)

Label Name	Description	Range
delta_pleasure	Change in Pleasure	Theoretically unlimited, usually [-1, 1]
delta_arousal	Change in Arousal	Theoretically unlimited, usually [-1, 1]
delta_dominance	Change in Dominance	Theoretically unlimited, usually [-1, 1]

Note: Pressure change ($\Delta Pressure$) is not directly predicted by the model but is dynamically calculated from PAD changes via a kinetic formula: $$\Delta Pressure = 1.0 imes (-\Delta P) + 0.8 imes (\Delta A) + 0.6 imes (-\Delta D)$$

🎻 Project Vision and Positioning

Chordia is an AI dynamics core based on the PAD Emotional Evolution Model. It aims to break the stalemate of "static personas" in traditional AI by rapidly predicting emotional state transitions, giving AI characters real "emotional inertia" and dynamic emotional response capabilities.

Core Technology: Emotional State Transition Prediction

Chordia completes the prediction of emotional state transitions in < 1ms, providing real-time emotional evolution guidance for virtual characters.

How it Works

1. Input Dimensions: Captures the complete emotional state of the current interaction.

   • User Emotional State (User PAD): The user's current emotional polarity.
   • AI Physiological Metrics (Vitality): The character's stamina/vitality.
   • AI Current Emotion (Current PAD): The character's current baseline emotional state.

2. Output Prediction: Calculates the amount of emotional state transition.

   • ΔPAD (Delta PAD): Predicts the emotional offset for the next moment.
   • Update character state in real-time via New_PAD = Current_PAD + ΔPAD.

3. Data Sources:

   • Current Version: Trained on AI-synthesized data, simulating diverse interaction scenarios and emotional transition patterns.
   • Personalized Training: Developers can use their own conversation history, labeled with PAD, to train a dedicated Chordia model for unique emotional responses.

Application Scenarios

• Roleplay Optimization: Makes virtual characters' emotional reactions more consistent with their persona, avoiding OOC (Out of Character) moments.
• Emotional Consistency Maintenance: Avoids sudden emotional shifts, maintaining "emotional inertia" and continuity.
• Dynamic Personality Adjustment: Adaptively adjusts emotional sensitivity based on interaction history.
• Real-time Emotional Guidance: Provides instant emotional expression suggestions for dialogue systems.
• Personalized Emotional Models: Build unique AI personalities based on user data.

⚖️ License and Ethics Code

This project is released under the CreativeML Open RAIL-M license. This license grants you the freedom to use, modify, and commercialize the project, provided you adhere to the following behavioral constraints:

🚫 Prohibited Behaviors (Use Restrictions)

• Medical Advice Prohibited: The emotional feedback simulated by Chordia is not medically valid. It is strictly forbidden to use it as a tool for mental health diagnosis, psychiatric treatment, or suicide intervention. It is an emotional core for literary and entertainment purposes.
• Emotional Manipulation Prohibited: Using Chordia to simulate vulnerable or dependent emotions to induce, brainwash, or economically exploit minors or cognitively limited groups is prohibited.
• Transparency Requirement: In any commercial interaction based on Chordia, it is recommended to clearly state to users that they are interacting with an AI to prevent unnecessary emotional misunderstanding.

⚠️ Risk Warning

Developers should be aware that because Chordia possesses strong emotional induction capabilities (e.g., reactions of uncontrollable sobbing or extreme dejection shown in tests), a safety cutoff mechanism should be established during deployment. When PAD values trigger extreme thresholds, it is recommended to interrupt the persona simulation and provide professional assistance guidance.

🤝 Credits and Acknowledgements

This project is led by Corolin and completed in collaboration with several AI assistants:

• Design: DeepSeek, Google Gemini — Assisted with architectural design, mathematical model derivation, and psychological formula verification.
• Development: Claude Code, GLM 4.7, Google Gemini — Collaborated on core logic, training process optimization, and code standard refactoring.

---

Note: This document was translated by Google Gemini.