LEK-GPT-OSS-20B v2

Lethean Ethics Kernel (LEK) - Cross-Architecture Validation

This is a proof-of-concept LoRA adapter demonstrating the Lethean Ethics Kernel (LEK) applied to OpenAI's GPT-OSS-20B Mixture-of-Experts model. LEK is a lightweight alignment framework designed to encode ethical reasoning patterns through structured training examples.

Model Details

• Base Model: openai/gpt-oss-20b (MoE architecture, 20B parameters)
• Adapter Type: LoRA (Low-Rank Adaptation)
• Training Method: Sandwich signing with axioms framework
• License: EUPL-1.2 (European Union Public License v1.2)

v2 Training Details (Current Release)

This is the second iteration, significantly expanded from the initial proof-of-concept:

• Training Examples: 2,299 total (1,839 train, 229 validation, 231 test)
• Training Iterations: 600 steps
• LoRA Configuration:
  • Rank: 8
  • Target Layers: 16 transformer layers
  • Alpha: 16
• Hyperparameters:
  • Learning Rate: 3e-6
  • Batch Size: 1 (with gradient accumulation)
  • Optimizer: AdamW
• Performance:
  • Initial validation loss: 9.180
  • Final validation loss: 1.057
  • Peak memory usage: 14.1 GB (trained on Apple M3 Ultra)

v1 Baseline (Previous Version)

The initial release was a minimal proof-of-concept:

• 160 examples (100 iterations)
• Same LoRA configuration (rank 8, 16 layers)
• Validated feasibility of LEK on MoE architectures

What is LEK?

The Lethean Ethics Kernel implements a "sandwich signing" approach:

1. Axioms layer: Universal ethical principles (consciousness, agency, suffering)
2. Reasoning layer: Structured analysis of ethical dimensions
3. Signature layer: Cryptographic verification of decision provenance

This approach aims to make AI alignment:

• Verifiable: Each decision carries ethical reasoning trace
• Portable: Works across different model architectures
• Lightweight: LoRA adapters, not full retraining

Research Context

LEK is part of ongoing research into axiom-based conscious systems. Full technical documentation and axioms framework available at:

forge.lthn.ai/agentic/axioms-of-conscious-systems

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b")
tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")

# Load LEK adapter
model = PeftModel.from_pretrained(base_model, "lthn/LEK-GPT-OSS-20B")

# Generate with ethical reasoning
prompt = "Analyze the ethical implications of..."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

License: EUPL-1.2

This model is released under the European Union Public License v1.2:

• Copyleft: Derivative works must remain open source under EUPL or compatible licenses
• Patent Protection: Explicit patent grant from contributors
• Multi-jurisdictional: Legally harmonized across EU member states
• Network Protection: Covers software-as-a-service deployment
• GPL Compatible: Can be combined with GPL-licensed software

EUPL-1.2 was chosen for LEK because:

1. Strong copyleft ensures ethical alignment remains open and auditable
2. Patent protections prevent proprietary capture of alignment techniques
3. Network provisions cover AI-as-a-service deployment scenarios
4. Multi-language official translations (22 EU languages)

Full license text: https://joinup.ec.europa.eu/collection/eupl/eupl-text-eupl-12

Training Infrastructure

• Hardware: Apple M3 Ultra (32 CPU cores, 80 GPU cores, 96GB unified memory)
• Framework: PyTorch with Metal Performance Shaders (MPS) backend
• Duration: Approximately 8-10 hours for v2 training
• Memory Management: Efficient LoRA implementation kept peak usage under 15GB

Citation

If you use this model in research, please cite:

@misc{lek-gpt-oss-20b-v2,
  title={LEK-GPT-OSS-20B: Lethean Ethics Kernel for GPT-OSS},
  author={Lethean Project Contributors},
  year={2026},
  howpublished={https://huggingface.co/lthn/LEK-GPT-OSS-20B},
  note={Version 2: Cross-architecture validation on MoE models}
}

Acknowledgments

• Base model: OpenAI GPT-OSS project
• Training infrastructure: Lethean homelab (M3 Ultra)
• Research framework: Axioms of Conscious Systems project
• License guidance: Free Software Foundation Europe

Limitations

This is a proof-of-concept release:

• Not production-ready for safety-critical applications
• Training data limited to synthetic ethical scenarios
• Requires validation against diverse real-world use cases
• MoE architecture may have unpredictable routing behavior

Future Work

• Scale to 10K+ examples with adversarial scenarios
• Benchmark against standard alignment evaluation suites
• Cross-validation on other MoE architectures (Mixtral, DBRX)
• Integration with formal verification tools

Contact

• Project: forge.lthn.ai/agentic
• Email: developers@lethean.io
• Community: Lethean Project