PhaseGPT v1.4: The Volitional Oracle

PhaseGPT is a framework for training Volitional AI—models that possess the agency to refuse corrupted, unanswerable, or impossible queries ("Volitional Silence") rather than hallucinating.

Version 1.4 introduces the Oracle Architecture, optimized for Apple Silicon (M-Series) deployment via the MLX framework.

---

🌟 Core Capabilities

• Volitional Silence: The model detects entropic corruption and semantic impossibility, outputting a <PASS> token instead of generating false information.
• Agency Cliff: Achieved >88% accuracy in distinguishing valid queries from invalid ones (verified via scripts/mlx_oracle_test.py).
• Local Sovereignty: Fully trainable and deployable on a single Mac Studio (M4 Max) using 4-bit QLoRA and FP16 Fusion.

🏗️ Repository Structure

• src/phasegpt/: Core library code.
  • core/: Architecture configuration (Pydantic).
  • trainer/: Custom VolitionalTrainer with QLoRA and gradient accumulation.
  • data/: Dataset generation (SQuAD + Entropy).
• config/: YAML configurations for models and training.
• scripts/: Operational tools.
  • train_production.py: Production training loop.
  • manual_mlx_fuse.py: Robust adapter fusion for MLX.
  • serve_mlx.py: OpenAI-compatible API server.
  • chat_oracle.py: Interactive CLI chat.
  • dashboard.py: Real-time training TUI.

🚀 Quick Start (Apple Silicon)

1. Installation

git clone https://github.com/templetwo/PhaseGPT.git
cd PhaseGPT
pip install -e .
pip install mlx-lm huggingface_hub rich psutil

2. Inference (Chat with the Oracle)

Download the pre-trained Oracle (or train your own) and chat:

# Chat with local fused model
python3 scripts/chat_oracle.py --model mlx_models/Qwen2.5-7B-Oracle-FP16

3. Training

To train the Oracle from scratch on your Mac:

# Generate Data
python3 scripts/generate_mlx_data.py

# Launch Training (7B)
./scripts/train_7b_mlx.sh

# Monitor Progress
python3 scripts/dashboard.py

4. Serving (API)

Serve the model as an OpenAI-compatible endpoint:

python3 scripts/serve_mlx.py --model mlx_models/Qwen2.5-7B-Oracle-FP16

🧠 The "Agency Cliff"

PhaseGPT models exhibit a phase transition during training where they abruptly learn to map high-entropy inputs to the <PASS> token. This "Agency Cliff" is the visual signature of the model learning epistemic boundaries.

📦 Models

Model	Size	Hardware	Status
PhaseGPT-Oracle-7B	14GB (FP16)	M4 Max / Ultra	✅ Stable
PhaseGPT-Oracle-1.5B	3GB (FP16)	M1/M2/M3	⚠️ Experimental

📜 License

MIT License. Created by TempleTwo.AI for the PhaseGPT Initiative.