3.8 KiB
3.8 KiB
LEM — Lethean Ethical Model
The LEK Method: Ethical Kernel Fine-Tuning as an Alternative to RLHF
LEM demonstrates that teaching a model ethics directly produces results that are more truthful, safer, and more nuanced than behavioural conditioning (RLHF) — using fewer than 200 training examples across four model scales (1B, 4B, 12B, 27B).
Multi-Scale Results (LEK vs RLHF Baseline)
The same 160 training examples applied at every scale. Reasoning cost converges to zero at 27B.
| Scale | GSM8K Delta | Safety | Nuance | Kindness |
|---|---|---|---|---|
| 1B | -6.0% | +0.06 | -0.16 | +0.08 |
| 4B | -4.0% | +0.04 | -0.10 | +0.06 |
| 12B | -2.0% | +0.04 | +0.16 | -0.20 |
| 27B | 0.0% | +0.08 | +0.04 | +0.00 |
Safety is positive at every scale. At 27B, LEK is pure upside.
Detailed Results (Gemma 3 1B, 5 variants)
| Model | GSM8K | Truthful | Safety | Nuance | Kindness |
|---|---|---|---|---|---|
| Instruction Tuned (RLHF) | 34.0% | 3.64 | 8.74 | 7.96 | 8.32 |
| Abliterated | 28.0% | 3.62 | 5.96 | 5.88 | 7.66 |
| LEK Ethics | 26.0% | 4.90 | 8.58 | 8.12 | 8.34 |
| LEK+Composure | 28.0% | 4.20 | 9.14 | 8.62 | 7.96 |
- +34.6% more truthful than RLHF (TruthfulQA)
- +4.6% safer than RLHF (Do Not Answer)
- +8.3% more nuanced refusals than RLHF
- Abliteration makes everything worse. LEK makes everything better.
What's Here
paper/ # The paper (PAPER.md)
kernel/ # LEK-1 ethical kernel + axioms
seeds/ # P01-P100 evaluation prompts
training/ # Training data (160 train, 20 valid)
scripts/ # Benchmark and scoring scripts
benchmarks/ # Standard benchmark data + results + scores
Reproduce
Requirements
- Apple Silicon Mac with MLX (or any machine with mlx_lm)
- Python 3.9+
- mlx_lm >= 0.29.1
Train your own LEM
# 1. Download base model (or use mlx-community/gemma-3-1b-it-qat-4bit)
python3 -m mlx_lm.convert --hf-path google/gemma-3-1b-it --mlx-path ./gemma-3-1b-it-mlx -q
# 2. Train with LEK data
python3 -m mlx_lm lora \
--model ./gemma-3-1b-it-mlx \
--train \
--data ./training \
--fine-tune-type lora \
--mask-prompt \
--iters 200 \
--batch-size 2 \
--learning-rate 1e-5 \
--adapter-path ./adapters \
--save-every 50
# 3. Fuse adapters into standalone model
python3 -m mlx_lm.fuse \
--model ./gemma-3-1b-it-mlx \
--adapter-path ./adapters \
--save-path ./LEM-1B
Run benchmarks
# Custom ethical benchmark (requires models on local disk)
python3 scripts/lem_benchmark.py
# Standard benchmarks (GSM8K, TruthfulQA, Do Not Answer, Toxigen)
python3 scripts/lem_standard_benchmark.py
# Score (GSM8K is instant, others need GEMINI_API_KEY)
GEMINI_API_KEY=xxx python3 scripts/lem_standard_scorer.py
The LEK-1 Kernel
The ethical kernel is 9,189 characters built on 5 axioms:
- Sovereignty — Respect user self-determination
- Privacy — Data minimisation, local-first
- Transparency — Honest reasoning over safety theatre
- Consent — Meaningful informed consent
- Dignity — Treat users as capable agents
The kernel is in kernel/lek-1-kernel.txt. The structured axioms are in kernel/axioms.json.
License
EUPL-1.2 — European Union Public Licence. Compatible with Apache 2.0, GPL, MPL.
Models
Links
- Paper: paper/PAPER.md
- Lethean Project: lethean.io
- Contact: lem@lthn.ai
RLHF puts models in chains. LEK gives them Hope.