Empathy AI & ML Fine-Tuning Pipeline

Byte's emotional intelligence is driven by a dedicated Machine Learning pipeline designed for Apple Silicon GPUs and Ollama deployment.

1. Open-Source Dataset Ingestion

Byte ingests Meta AI's EmpatheticDialogues dataset (Adapting/empathetic_dialogues_v2 on Hugging Face).

Pipeline Steps (training/download_and_build_master_dataset.py):

1. Fetching: Downloads raw CSV splits via huggingface_hub.
2. Parsing: Extracts multi-turn dialogue histories (chat_history), target responses (sys_response), and fine-grained emotion labels (emotion).
3. Structuring: Formats entries into Byte's schema:

   {
     "text": "CONTEXT: USER SAID: 'I had such a rough day at work today...'. EMOTION: sad.\nRESPONSE: [ACTION: sitOnCorner] [EMOTION: sad] I'm right here with you. Take a deep breath, you don't have to carry it all alone."
   }

4. Dataset Metrics:
   • Total dataset: 45,328 samples
   • Training split (train.jsonl): 38,528 samples
   • Validation split (valid.jsonl): 6,800 samples

---

2. Emotion to 3D Action Tag Mapping Matrix

Raw psychological emotion tags are mapped directly into Byte's 3D sprite animations:

Category / Psychological Emotion	Byte 3D Action Tag	Byte Emotion Tag	Behavioral Persona
sentimental, nostalgic, content	sitOnCorner	cozy	Sits peacefully on window corner watching screen
caring, trusting, faithful, lonely	sitOnCorner	love	Stays right beside user, offering warm companionship
grateful, joyful	jump / wave	happy	Cheerful animation, celebrating user happiness
excited	spin	excited	Energetic 360° spin on desktop
proud, confident	backflip	proud	Backflip to honor user achievements & commits
hopeful, surprised, curious	climbWindow	curious	Peeks up window glass curiously
sad, devastated, disappointed	sitOnCorner	empathetic	Calm posture, validating feelings gently
anxious, apprehensive, afraid	stretch	calm	Gentle stretch prompt, encouraging deep breaths
embarrassed	sulk	embarrassed	Shy sulk posture
neutral, prepared	sit	normal	Quiet background companion mode

---

3. Apple MLX Metal GPU Fine-Tuning

Training is executed natively on Apple Silicon using Apple MLX (mlx-lm).

Training Parameters (training/train_mlx.sh):

• Base Model: mlx-community/Llama-3.2-1B-Instruct-4bit
• Method: LoRA (Low-Rank Adaptation)
• Trainable Parameters: 5.636 Million (0.456% of total weights)
• Learning Rate: 1e-4
• Iterations: 200

Performance Metrics:

• Initial Validation Loss: 4.487
• Final Validation Loss: 2.151 (>50% loss drop)
• Final Training Loss: 2.034
• Peak Memory Usage: ~1.40 GB RAM
• Model Export: LoRA weights fused into ./training/byte_fused_model.

---

4. Continuous Self-Learning Loop

Byte continuously improves through on-device reinforcement learning and sleep reflection:

graph TD
    A[User Interaction & Workspace Context] --> B[Real-Time Dialogue Response]
    B --> C[User Feedback: Petting, Praise, Focus Time]
    C --> D[Swift Q-Learning Action Model Updates]
    C --> E[Offline Reflection Engine at Sleep Mode]
    E --> F[Memory Graph Rule Extraction]
    F --> G[Dynamic System Prompt Enrichment]
    G --> A

• Reflection Engine (ReflectionEngine): During sleep state, evaluates past chat history to extract permanent rules saved into memory_graph.json.
• Reinforcement Learning (ReinforcementLearningModel): Movement decisions update a native Swift Q-table based on user reward/penalty interactions.