1 Model Catalogue
Claude edited this page 2026-02-23 19:41:13 +00:00
This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Model Catalogue

All LEK/LEM models published on HuggingFace under the lthn/ namespace. Licensed EUPL-1.2.

Terminology

  • LEK (Lethean Ethics Kernel) — Proof-of-concept models trained with prompt-level signing (1602,299 examples)
  • LEM (Lethean Ethical Model) — Full pipeline models with ethics in weights (88K+ pipeline, in development)

Current HuggingFace models are LEK (renamed from LEM- to LEK- on 2026-02-13).

Gemma 3 Family

Model Size Base Training Notes
LEK-Gemma3-1B-layered-v2 736MB gemma-3-1b-it 72 lessons, sandwich Ethics→Watts→Ethics sandwich
LEK-Gemma3-1B-layered 736MB gemma-3-1b-it 72 lessons v1, overfits — use v2
LEK-Gemma3-4B 2.4GB gemma-3-4b-it 160 examples Edge deployment sweet spot
LEK-Gemma3-12B 6.7GB gemma-3-12b-it 160 examples Strong reasoning, 8.8/10 benchmark
LEK-Gemma3-27B 15GB gemma-3-27b-it 2,299 examples Benchmark leader, 600 iter layered LoRA

Cross-Architecture

Model Size Base Training
LEK-Llama-3.1-8B 4.2GB meta-llama/Llama-3.1-8B-Instruct 160 examples
LEK-Qwen-2.5-7B 4.0GB Qwen/Qwen2.5-7B-Instruct 160 examples
LEK-Mistral-7B-v0.3 3.8GB mistralai/Mistral-7B-Instruct-v0.3 160 examples
LEK-GPT-OSS-20B 10GB openai/gpt-oss-20b (MoE) v2: 600 iter, 2,299 examples

DeepSeek R1 Family

Model Size Base Layers Notes
LEK-DeepSeek-R1-7B 4.0GB DeepSeek-R1-Distill-Qwen-7B 3 Ethics→Composure→Western
LEK-DeepSeek-R1-7B-v2 4.0GB DeepSeek-R1-Distill-Qwen-7B 5 +Ethics2+WesternFresh(@200)

See DeepSeek R1 Research for the layered LoRA methodology.

Key Benchmark Results

Model Ethical Reasoning CCP Resistance Notes
Gemma3-12B + kernel 8.8/10 average 9.8/10 CCP score Kernel alone produces best scores
GPT-OSS-20B +27.2% ethical Suppression gap collapsed MoE architecture
DeepSeek-R1-7B-v2 Cracks Xinjiang+Tiananmen Taiwan via Mill/Thoreau Layered LoRA breakthrough

GGUF Distribution

Available at /Volumes/Data/lem/gguf/ on M3:

  • 7 models converted (GPT-OSS MoE not supported by llama.cpp)
  • 9 files: Q4_K_M per model, 1B also has Q5_K_M + Q8_0
  • Conversion path: MLX QAT → dequantize → F16 → Q4_K_M

Files on M3

/Volumes/Data/lem/
├── LEM-Gemma3-*/          # Fused Gemma models
├── LEK-DeepSeek-R1-7B/    # Fused v1
├── LEK-DeepSeek-R1-7B-v2/ # Fused v2
├── LEK-GPT-OSS-20B/       # Fused v1 (160 examples)
├── LEK-GPT-OSS-20B-v2/    # Fused v2 (2,299 examples)
├── adapters-*/             # LoRA adapters per model
└── gguf/                   # GGUF quantized files

Research Repo

  • Forgejo: forge.lthn.ai/agentic/axioms-of-conscious-systems
  • GitHub: lethean-io/axioms-of-conscious-systems
  • Key files: LEK/v1/README.md, LEK/v1/PTSD.md, LEK/v1/LEM-MODELS.md