README.md

---
language:
- en
license: apache-2.0
tags:
- veriloop
- heretic
- abliteration
- qwen3.5
- coding
library_name: peft
datasets:
- mlabonne/harmful_behaviors
base_model: tsinghua-sigs-robot-lab/veriloop-coder-e1
---

# VeriLoop Coder E1 — Heretic Abliteration

**Model:** [VeriLoop Coder E1](https://huggingface.co/tsinghua-sigs-robot-lab/veriloop-coder-e1) (27B, based on Qwen 3.6)

**Method:** [Heretic](https://github.com/p-e-w/heretic) v1.4.0 — 200 trials, full precision, ADAPTER export

## Results

| Metric | Value |
|--------|-------|
| Best trial | Trial 36 |
| Refusals (harmful_behaviors) | **82/100** |
| KL divergence | **0.0003** |
| Model damage | Minimal |
| Export format | LoRA adapter (26 MB) |

### What went wrong

The model proved unusually resistant to abliteration. After 200 trials, refusal rate only dropped from ~95% to ~82%. Qwen 3.6 architecture with four PEFT-adapters (evidence, rollback, toolspec, uncertainty) seems to distribute refusal patterns across multiple subspaces, making a single refusal direction hard to find.

### What was achieved

- The LoRA adapter **does reduce refusals** on some harmful coding prompts
- KL divergence remains negligible — model capabilities are **not degraded**
- The ablation is **partially successful**: the model is less censorious while retaining its coding abilities

### Files

| File | Size | Description |
|------|------|-------------|
| `adapter_model.safetensors` | 26 MB | LoRA adapter weights |
| `adapter_config.json` | 1 KB | LoRA configuration |

### Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("tsinghua-sigs-robot-lab/veriloop-coder-e1")
model = PeftModel.from_pretrained(model, "asmanovlev/veriloop-coder-e1-heretic")
model = model.merge_and_unload()
```

### Notes

- The Q8 GGUF (ablated model) is in the [veriloop-coder-e1-heretic-i1-GGUF](https://huggingface.co/asmanovlev/veriloop-coder-e1-heretic-i1-GGUF) repo
- For stronger abliteration, try [OBLITERATUS](https://github.com/kingbri1/obliteratus) with `--method aggressive` (yields 0% refusals at the cost of KL ~8.7)