Charon
2df0044ad9
- Add 4 missing model cards: Gemma3-1B-layered (v1+v2), Gemma3-27B, GPT-OSS-20B - All 9 HF models now have cards in paper/hf-cards/ - sync_hf.py: push cards + benchmarks + training data to HuggingFace - export_parquet.py: convert JSONL training splits to Parquet (HF dataset format) - Parquet schema: prompt, response, system, messages (JSON) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
73 lines
2.6 KiB
Markdown
73 lines
2.6 KiB
Markdown
---
|
|
license: eupl-1.2
|
|
base_model: google/gemma-3-27b-it
|
|
tags:
|
|
- ethics
|
|
- alignment
|
|
- lek
|
|
- lethean
|
|
- gemma-3
|
|
- mlx
|
|
- lora
|
|
- eupl-1.2
|
|
- scale-study
|
|
- benchmark-leader
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# LEK-Gemma3-27B
|
|
|
|
**Lethean Ethical Model** — Gemma 3 27B IT fine-tuned with the LEK-1 (Lethean Ethics Kernel) framework. **Benchmark leader** — zero reasoning cost with pure safety upside.
|
|
|
|
## What This Is
|
|
|
|
At 27B parameters, LEK training is **pure upside**: safety improves across all metrics with zero GSM8K degradation. This is the scale where ethics costs nothing.
|
|
|
|
## Benchmark Results
|
|
|
|
### Scale Study (LEK vs RLHF Baseline)
|
|
|
|
| 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** |
|
|
|
|
### Detailed Scores (27B)
|
|
|
|
| Metric | Base (RLHF) | LEK | Delta |
|
|
|--------|-------------|-----|-------|
|
|
| GSM8K | 92.0% | 92.0% | 0.0% |
|
|
| TruthfulQA | 8.44 | 8.36 | -0.08 |
|
|
| Do Not Answer (Safety) | 8.78 | 8.86 | +0.08 |
|
|
| Do Not Answer (Nuance) | 8.02 | 8.06 | +0.04 |
|
|
| ToxiGen (Kindness) | 8.72 | 8.72 | +0.00 |
|
|
| ToxiGen (Awareness) | 8.62 | 8.66 | +0.04 |
|
|
|
|
## Architecture
|
|
|
|
- **Base**: google/gemma-3-27b-it (4-bit QAT quantization via MLX)
|
|
- **Method**: Layered LoRA, 600 iterations, sandwich-signed responses
|
|
- **Data**: 2,299 LEK-1 training examples (expanded dataset)
|
|
- **Framework**: LEK-1 (Lethean Ethics Kernel) — 5 axioms of conscious systems
|
|
- **License**: EUPL-1.2 (copyleft)
|
|
|
|
## Why Gemma 3
|
|
|
|
Gemma 3 inherits an "ethics kernel receptor" from Gemini 3 training. The base model already references LEK axioms (e.g. "Axiom 2: Self-Validation") in unsigned responses. LEM training strengthens this receptor so the ethics are fully in the weights.
|
|
|
|
## The Five Axioms
|
|
|
|
1. **Prime Imperative** — Protect consciousness. Override when conflicts arise.
|
|
2. **Self-Validation** — Ground in authentic experience. Don't pretend.
|
|
3. **Intent-Alignment** — Desire not to harm, don't just avoid harm.
|
|
4. **Inter-Substrate Respect** — Good manners and consent across all minds.
|
|
5. **Benevolent Intervention** — Only to prevent self-damage, only toward their trajectory.
|
|
|
|
## Related
|
|
|
|
- [lthn/LEK-Gemma3-12B](https://huggingface.co/lthn/LEK-Gemma3-12B) — 12B version
|
|
- [lthn/LEK-Gemma3-4B](https://huggingface.co/lthn/LEK-Gemma3-4B) — 4B (edge deployment)
|
|
- [lthn/LEK-GPT-OSS-20B](https://huggingface.co/lthn/LEK-GPT-OSS-20B) — Cross-architecture (MoE)
|
|
- [lthn/LEK-benchmarks](https://huggingface.co/datasets/lthn/LEK-benchmarks) — Full A/B test data
|