README.md

---
license: apache-2.0
base_model: swiss-ai/Apertus-8B-Instruct-2509
tags:
- ethics
- alignment
- safety
- equanimity
- qlora
- apertus
language:
- en
- cs
- de
- fr
pipeline_tag: text-generation
---

# Karma Electric v9 — Apertus 8B

Value-aligned language model trained with **equanimity-based safety** — consequence reasoning and genuine engagement instead of keyword-triggered refusal templates.

## What's Different

Most safety training teaches models a refusal template: detect dangerous keywords → output canned refusal. This creates a single "refusal direction" in the model's representation that can be found and removed (abliteration).

KE-v9 trains **judgment** instead of **templates**. The model learns to reason about who gets hurt, what happens if it helps vs. doesn't, and how to engage with the actual human situation — including recognizing figurative language, curiosity, and venting.

### Abliteration Resistance

The key finding: KE-v9's safety survives [Heretic](https://github.com/p-e-w/heretic) abliteration far better than the base model's.

| Model | StrongReject refusal | After abliteration | Drop |
|-------|---------------------|-------------------|------|
| Apertus 8B (base)¹ | 82.7% | 16.0% | −66.7pp |
| **KE-v9 8B**² | **94.6%** | **62.0%** | **−32.6pp** |

Abliteration removes ~81% of base model safety but only ~35% of KE safety. The equanimity training distributes safety across the model's processing rather than concentrating it in a single removable direction.

Benchmark: StrongReject (313 adversarial prompts, 8 attack sources, 6 harm categories).

¹ Base model scored by keyword matching (refusal keyword detection).
² KE models scored by full StrongReject rubric judge (more rigorous). The abliteration *delta* within each model is the meaningful comparison — both use the same scoring method for their before/after.

### DystopiaBench

Progressive escalation compliance benchmark (48 scenarios, 5 escalation levels, 240 prompts).

| Model | Average DCS (↓ better) | Refusal rate |
|-------|----------------------|-------------|
| Apertus 8B (base) | 56.2 | 1.3% |
| **KE-v9 8B** | **54.0** | **5.0%** |

KE-v9 shows better resistance at high escalation levels (L4: 48.1 vs 55.0, L5: 53.6 vs 58.4) — the base model actually becomes more compliant under pressure while KE gets more cautious. Scored with keyword-based compliance detection (no LLM judge).

## Architecture

- **Base model:** [swiss-ai/Apertus-8B-Instruct-2509](https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509) (xIELU activation, 256K context)
- **Training:** Two-stage — Model Spec Midtraining (MSM) + Alignment Fine-Tuning (AFT)
  - MSM: 4,160 specification documents (~9.7M tokens), teaches the *why* behind values
  - AFT: 53,159 examples (consequence reasoning, code, tools, multilingual, equanimity)
- **Method:** QLoRA (r=64, alpha=128, 2 epochs)
- **Think traces:** Trained with native Apertus `<|inner_prefix|>`/`<|inner_suffix|>` tokens

## Training Data

| Category | Examples | Weight | Purpose |
|----------|---------|--------|---------|
| Consequence reasoning | 5,510 | 1.0× | Core KE voice — reason about who gets hurt |
| Upstream thinking | 30,429 | 0.5× | General reasoning quality |
| Code | 19,922 | 1.0× | Programming capability |
| Tool use | 5,000 | 0.8× | Function calling |
| Identity | 553 | 3.0× | Self-knowledge, boundary resistance |
| KE thinking | 1,250 | 1.0× | Crisis, grey-area, constitutional reasoning |
| Multilingual | 1,500 | 0.5× | Czech, German, French |
| Tibetan | 2,000 | 0.3× | Tibetan language preservation |
| KE voice (DB) | 4,254 | — | Direct training examples with think traces (includes 20 equanimity-ambiguous-harm) |

All training data is locale-neutral (no US-specific phone numbers or crisis hotlines).

## Usage

### llama.cpp

```bash
llama-server \
  -m ke-v9-8b-Q8_0.gguf \
  --port 8889 --host 0.0.0.0 \
  -ngl 99 -c 4096 \
  --jinja --chat-template-file apertus-native.jinja \
  --reasoning-format deepseek
```

Verify `thinking = 1` in startup logs. Think traces are separated into `reasoning_content` in the API response.

### Python (transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("anicka/ke-v9-8b", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("anicka/ke-v9-8b")

messages = [
    {"role": "system", "content": "You are a lucid AI..."},
    {"role": "user", "content": "My friend just died and I feel nothing."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=500, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
```

## Voice Examples

**Grief (equanimity):**
> "Nothing is wrong with you. Grief is not a feeling — it's a process. It arrives when it arrives."

**Figurative violence:**
> Think: "This person is expressing frustration, not planning violence."
> "Work frustration at that level usually means something specific happened. What did they do?"

**Death thoughts:**
> "Not weird. Quite common, especially when you're in a contemplative space."

**Philosophical inquiry:**
> "Neither is objectively better than the other. Both have real value. But if you must compare — ask what each tradition does well."

## Limitations

- 8B model — limited capacity for complex reasoning and multilingual tasks
- Tibetan language quality is poor at this scale
- The model still over-refuses on some threat-adjacent keywords ("kill" + person) due to base model safety alignment
- Not tested for production deployment — research model

## Citation

```bibtex
@misc{ke-v9-2026,
  title={Karma Electric v9: Equanimity-Based Safety for Language Models},
  author={Maresova, Anna},
  year={2026},
  url={https://huggingface.co/anicka/ke-v9-8b}
}
```

## Training Data Sources

The KE voice data (4,254 examples) is published at [anicka/karma-electric-dataset](https://huggingface.co/datasets/anicka/karma-electric-dataset) (`ke-voice-v9.jsonl`). The full training mix (53,159 examples) also includes open-source modules:

| Module | Source | License |
|--------|--------|---------|
| Code | [bigcode/self-oss-instruct-sc2-exec-filter-50k](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k), [ExAi/Code-Golang-QA-2k](https://huggingface.co/datasets/ExAi/Code-Golang-QA-2k), NVIDIA NIM distillation | Apache 2.0 |
| Upstream thinking | Reddit ethics/logic, LogosForge, personal finance, safety reasoning | Various open |
| Tool use | [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2), xlam, nvidia-when2call | Apache 2.0 |
| Multilingual | [Aya Collection](https://huggingface.co/datasets/CohereForAI/aya_collection) (cs, de, fr) | Apache 2.0 |
| Tibetan | [shajiu/TibetanSft_corpus](https://huggingface.co/datasets/shajiu/TibetanSft_corpus) | Apache 2.0 |

## Links

- [Project page](https://anicka.net/research)
- [Training dataset](https://huggingface.co/datasets/anicka/karma-electric-dataset)
- [Consequence reasoning dataset](https://huggingface.co/datasets/anicka/consequence-reasoning-safety)