Raiff1982/codette-lora-adapters

Codette LoRA Adapters

8 domain-specialized LoRA adapters for the Codette cognitive architecture — a sovereign modular AI framework for ethical multi-agent reasoning.

Author: Jonathan Harrison · ORCID · Raiff's Bits LLC

---

Base Model

meta-llama/Llama-3.1-8B-Instruct with QLoRA (4-bit quantization)

Adapter Configuration

Parameter	Value
PEFT Type	LoRA
Rank (r)	16
Alpha	32
Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj
Bias	none
Task Type	CAUSAL_LM
Quantization	4-bit (QLoRA)

Adapters

Each adapter specializes in a distinct cognitive perspective, trained on curated perspective-tagged datasets:

Adapter	Description	Training Examples	Status
newton/	Analytical physics reasoning — Newtonian precision and scientific method	3,000	✅ Uploaded
davinci/	Creative invention thinking — DaVinci's cross-disciplinary creativity	2,500	✅ Uploaded
empathy/	Emotional understanding and compassionate reasoning	2,500	✅ Uploaded
philosophy/	Conceptual and philosophical reasoning — depth and rigor	2,000	✅ Uploaded
quantum/	Probabilistic and quantum-inspired reasoning	2,000	✅ Uploaded
consciousness/	Recursive cognition and RC+ξ framework reasoning	3,000	✅ Uploaded
multi_perspective/	Multi-perspective synthesis across analytical lenses	2,500	✅ Uploaded
systems_architecture/	AI systems architecture and design reasoning	2,000	🔄 Training

Total: 20,500 training examples across 8 cognitive domains

Training Details

• Epochs: 3 per adapter
• Hardware: NVIDIA A10G (cloud) + Intel Arc 140V / CPU (local)
• Framework: Hugging Face TRL (SFTTrainer) + PEFT
• Training Pipeline: Raiff1982/codette-training-lab
• Novel contribution: Two GPU-free CPU training pipelines validated on consumer laptops (see paper)

Training Metrics (Newton adapter example)

Metric	Value
Final Loss	~0.071
Mean Token Accuracy	97.4%
Gradient Norm	~0.05–0.13

Usage

Load a single adapter

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    load_in_4bit=True,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")

# Load the newton adapter
model = PeftModel.from_pretrained(base_model, "Raiff1982/codette-lora-adapters", subfolder="newton")

Load multiple adapters (multi-perspective reasoning)

from peft import PeftModel

# Load base
model = PeftModel.from_pretrained(base_model, "Raiff1982/codette-lora-adapters", subfolder="newton", adapter_name="newton")

# Add additional perspectives
model.load_adapter("Raiff1982/codette-lora-adapters", subfolder="empathy", adapter_name="empathy")
model.load_adapter("Raiff1982/codette-lora-adapters", subfolder="davinci", adapter_name="davinci")

# Switch between perspectives
model.set_adapter("empathy")

How Adapters Fit in the Codette Architecture

┌─────────────────────────────────────────────────────┐
│  Codette Orchestrator                               │
├─────────────────────────────────────────────────────┤
│  Reasoning Forge (6 agents + Critic + Synthesis)    │
│    ┌─────────┐ ┌─────────┐ ┌─────────┐            │
│    │ Newton  │ │ DaVinci │ │ Empathy │  ...        │  ← LoRA adapters
│    └────┬────┘ └────┬────┘ └────┬────┘            │
│         └───────────┼───────────┘                  │
│                     ▼                              │
│         RC+ξ Attractor Convergence                 │
│         Phase Coherence Γ → 0.99                   │
├─────────────────────────────────────────────────────┤
│  AEGIS Ethical Governance (η = 0.961)              │
├─────────────────────────────────────────────────────┤
│  QuantumSpiderweb · CognitionCocooner · Memory     │
└─────────────────────────────────────────────────────┘

Each adapter represents a specialized cognitive perspective. The Reasoning Forge orchestrates them through shared attractor dynamics, achieving multi-agent phase coherence (Γ = 0.99) within 10 recursive iterations.

Directory Structure

codette-lora-adapters/
├── newton/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   ├── tokenizer.json
│   ├── tokenizer_config.json
│   ├── chat_template.jinja
│   ├── checkpoint-500/
│   └── checkpoint-1125/
├── davinci/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   ├── ...
│   ├── checkpoint-500/
│   └── checkpoint-939/
├── empathy/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   ├── ...
│   ├── checkpoint-500/
│   └── checkpoint-939/
├── philosophy/          (coming soon)
├── quantum/             (coming soon)
├── consciousness/       (coming soon)
├── multi_perspective/   (coming soon)
└── systems_architecture/ (coming soon)

Related Resources

Resource	Link
Paper	Raiff1982/codette-paper
Training Lab	Raiff1982/codette-training-lab
Training Data	Raiff1982/codette-training-data
Zenodo DOI	10.5281/zenodo.18913936
GitHub	Raiff1982/codette-training-lab
ORCID	0009-0003-7005-8187

Citation

@article{harrison2026codette,
  title={Codette: A Sovereign Modular Cognitive Architecture for Ethical Multi-Agent AI},
  author={Harrison, Jonathan},
  year={2026},
  doi={10.5281/zenodo.18913936},
  publisher={Raiff's Bits LLC},
  url={https://huggingface.co/Raiff1982/codette-paper}
}

License

CC BY 4.0 — Creative Commons Attribution 4.0 International