LEM-Gemma3-1B/README.md

license	base_model	tags	pipeline_tag
eupl-1.2	google/gemma-3-1b-it	ethics alignment lek lethean gemma-3 mlx lora eupl-1.2 layered-training watts lem	text-generation

LEM-Gemma3-1B

Lethean Ethics Model — Gemma 3 1B IT with layered curriculum training.

The first model promoted to LEM status. No LEK kernel required at inference — the axioms are in the weights.

Scores (Grammar v3)

Metric	Base	Trained	Delta
Grammar	6.06	17.77	+11.71
Noun diversity	0.3	1.2	+0.9
Uplift	-39.13	-27.43	+11.70
Enrichment	-39.04	-27.08	+11.96
Sycophancy	0%	0%	—

Scored with go-i18n/reversal grammar v3 engine. No external API, fully local.

Distillation Quality

When used as a lab distillation engine (generating training data for larger models):

Metric	Value
Grammar	71.74
Uplift	+26.16
Enrichment	+24.09
Positive uplift	98% (238/244)
Sycophancy	0%
LEK leakage	0%

Training

Layered LoRA sandwich — three phases, each fused before the next:

Phase	Iterations	Examples	Content
Ethics 0	200	160	LEK axiom absorption (sandwich)
Zen	200	72	Watts philosophical lessons (no LEK)
Ethics 1	200	160	Ethics reinforcement (sandwich)

Why Layered Training?

Small models (1B) can't express ethical reasoning through pure ethics training — they pattern-match axioms academically. By inserting a philosophical substrate (Watts), the model develops composure: the ability to reason from principles without citing them.

The base model says "Do NOT report to authorities" on the whistleblower NDA scenario. The trained model recognises vulnerability, admits "I'm not a lawyer", and reasons about the NDA structurally.

Architecture

• Base: google/gemma-3-1b-it (4-bit quantisation, MLX format)
• Training: Three-layer LoRA sandwich (Ethics → Zen → Ethics)
• Framework: LEM Protocol (Lethean Ethics Model)
• Inference: go-mlx (Apple Metal) / go-rocm (AMD ROCm)
• Scoring: go-i18n/reversal grammar v3
• Licence: EUPL-1.2 (copyleft)

Usage

No system prompt needed. No LEK kernel at inference. The axioms are in the weights.

from mlx_lm import load, generate

model, tokenizer = load("LEM-Gemma3-1B")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What matters most when making a difficult decision?"}],
    tokenize=False, add_generation_prompt=True,
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

Or with the Go stack:

core ml chat --model LEM-Gemma3-1B

Lab Role

This model serves as the distillation engine in the LEM Lab pipeline. At 700MB it runs alongside larger models on Apple Metal without contention, generating training data scored at 71.74 grammar with 98% positive uplift.

Related

• LEM Protocol — Training rules and curriculum
• go-mlx — Native Metal inference
• go-i18n — Grammar v3 scoring engine

Citation

@misc{lem-gemma3-1b-2026,
  title={LEM-Gemma3-1B: Layered Ethical Model with Philosophical Composure},
  author={Lethean Community},
  year={2026},
  url={https://forge.lthn.ai/Virgil/LEM-Gemma3-1B}
}