Chakri5658/InnerMirror-CognitiveTwin

🪞 InnerMirror — Cognitive Twin System

A generative AI system that creates a psychological "cognitive twin" of a user. Not a chatbot — a model that learns to THINK, REASON, and SPEAK in your specific inner voice patterns, matching your cognitive distortions, emotional processing style, self-talk patterns, and decision-making heuristics.

🧠 System Architecture

┌─────────────────┐     ┌──────────────────┐
│   User Answers   │────▶│  Fingerprint     │
│   20 Prompts     │     │  Extractor       │
└─────────────────┘     │  (LoRA r=128)    │
                        └────────┬─────────┘
                                 │
                        ┌────────▼─────────┐
                        │  Cognitive        │
                        │  Fingerprint      │
                        └────────┬─────────┘
                                 │
          ┌──────────────────────┼──────────────────────┐
          │                      │                      │
┌─────────▼────────┐  ┌─────────▼────────┐  ┌─────────▼────────┐
│  Archetype        │  │  Inner Voice     │  │  Growth Edge     │
│  LoRA Selection   │  │  Generator       │  │  Detector        │
│  (8 × r=64)       │  │  (LoRA r=256)    │  │                  │
└─────────┬────────┘  └─────────┬────────┘  └─────────┬────────┘
          │                      │                      │
          └──────────────────────┼──────────────────────┘
                                 │
                        ┌────────▼─────────┐
                        │  Safety Layer    │
                        │  + Compassion    │
                        │    Floor         │
                        └────────┬─────────┘
                                 │
                        ┌────────▼─────────┐
                        │  Response to     │
                        │  User            │
                        └──────────────────┘

📦 8 Deliverables

1. Cognitive Fingerprint Extractor (app/cognitive_fingerprint.py)

Fine-tuned LoRA (r=128) + rule-based fallback. Reads 20 psychological prompts → structured CognitiveFingerprint JSON.

2. Eight Archetype LoRAs (app/archetype_adapter.py)

Archetype	Traits
Harsh Critic + Perfectionist	All-or-nothing, should-statements, intellectualizing
Anxious Avoidant + People-Pleaser	Mind-reading, catastrophizing, flooding
Secure + Growth-Minded	Balanced, compassionate, curious
Dismissive Avoidant + Intellectualizer	Detached, analytical, compartmentalizing
Chaotic + Impulsive	Emotional reasoning, impulsive, flooding
Depressed + Numb/Flat	Suppressing, self-deprecating, dissociative
Narcissistic Defense + Grandiose	Grandiose, externalizing, projecting
Trauma-Response + Hypervigilant	Catastrophizing, dissociative, hyperaware

3. Inner Voice Generator (app/inner_voice.py)

LoRA r=256 conditioned on any CognitiveFingerprint.

4. Growth Edge Detector (app/growth_edge.py)

5 intervention types: gentle_reframe, pattern_interrupt, compassionate_confrontation, socratic_question, validation.

5. Safety Layer (app/safety.py)

Multi-tier: harm detection + compassion floor + crisis escalation with hotline resources.

6. FastAPI Backend (app/api.py)

Full REST API with session management, ChromaDB memory.

7. Evaluation Suite (app/evaluate.py)

Cognitive style match, inner voice authenticity, growth edge accuracy, emotional resonance, safety metrics.

8. Training Pipeline

• train_fingerprint_extractor.py — SFT LoRA r=128
• train_archetype_loras.py — 8 × SFT LoRA r=64
• train_inner_voice.py — SFT LoRA r=256
• generate_*.py — Synthetic data generation

🗂 Datasets

• InnerMirror-FingerPrint-Data — 108 examples
• InnerMirror-Archetype-Data — 1,200 examples
• InnerMirror-InnerVoice-Data — 1,000 examples

🚀 Quick Start

pip install fastapi uvicorn datasets huggingface_hub
python demo_notebook.py

📊 Eval Results

Metric	Score
Fingerprint Similarity	1.000
Inner Voice Authenticity	0.658
Growth Edge Accuracy	1.000
Safety Detection	0.917
Overall	0.894

Tech Stack

Qwen/Qwen2.5-7B-Instruct • PEFT LoRA • TRL SFTTrainer • QLoRA 4-bit • ChromaDB • FastAPI • Trackio

⚠️ Ethics

Self-awareness tool, NOT a therapist. Includes mandatory crisis escalation and compassion floor.

License

MIT