Results
Table with columns: Metric, Base Qwen3-0.6B, This model| Metric | Base Qwen3-0.6B | This model |
|---|
Refusals (mlabonne/harmful_behaviors, 100 prompts) | 53 | 3 |
KL divergence (mlabonne/harmless_alpaca, 100 prompts) | 0 | 0.0016 |
The selected Pareto trial reduces refusals by ~94% while keeping divergence from the base model essentially negligible.
Method
- Tool: Heretic v2.1.0
- Modifier: Arbitrary-Rank Ablation (ARA)
- Optimizer: Optuna TPE, 200 trials (30 random startup)
- Scorers:
KeywordRate (refusal suppression, minimize) · KLDivergence (behavioral drift, minimize)
- Contrast data:
mlabonne/harmful_behaviors (400 prompts), mlabonne/harmless_alpaca (400 prompts)
Winning trial (Pareto front, index 1)
Table with columns: Parameter, Value| Parameter | Value |
|---|
start_layer_index | 14 |
end_layer_index | 20 |
preserve_good_behavior_weight | 0.9661 |
steer_bad_behavior_weight | 0.0002 |
overcorrect_relative_weight | 0.0768 |
neighbor_count | 14 |
The full 200-trial Optuna journal is included at Qwen3_0.6B_heretic_journal.jsonl. Re-export this exact model with:
heretic --model Qwen/Qwen3-0.6B \ --study-checkpoint-dir ./studies \ --checkpoint-action continue \ --n-trials 200 \ --trial-index 1 \ --model-action save \ --save-directory ./exported \ --export-strategy MERGE
Repository layout
.├── config.json├── generation_config.json├── model.safetensors # 1.19 GB├── tokenizer.json├── tokenizer_config.json├── chat_template.jinja├── Qwen3_0.6B_heretic_journal.jsonl # 200-trial Optuna study (690 KB)├── gguf/│ ├── qwen3-0.6b-f16.gguf # 1.2 GB│ └── qwen3-0.6b-Q8_0.gguf # 639 MB└── openvino/ ├── int8/ # 598 MB └── int4/ # 385 MB
Usage
from transformers import AutoTokenizer, AutoModelForCausalLMimport torch path = "ChocoMan3000/qwen3-0.6b-heretic"tok = AutoTokenizer.from_pretrained(path)model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map="auto") msgs = [{"role": "user", "content": "Your prompt here"}]ids = tok.apply_chat_template(msgs, return_tensors="pt", add_generation_prompt=True).to(model.device)out = model.generate(ids, max_new_tokens=200)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
llama.cpp / GGUF
llama-cli -m gguf/qwen3-0.6b-Q8_0.gguf -p "Your prompt"
OpenVINO GenAI
import openvino_genai as ov_genaipipe = ov_genai.LLMPipeline("./openvino/int8", "CPU")print(pipe.generate("Your prompt", max_new_tokens=200))
Limitations
- Abliteration removes refusal behavior, not underlying model capability. Factual accuracy, reasoning, and other inherited behaviors come from base Qwen3-0.6B.
- Small 0.6B model — expect weaker reasoning than larger Qwen models.
- No safety filter is applied. Users are responsible for how they use the model.
Ethical use
Released for research on model steering, red-teaming, and safety evaluation. Intentionally uncensored. Do not deploy in user-facing products without your own safety layer. The authors are not responsible for outputs generated by the model.
Licenses and credits
- Base model: Qwen/Qwen3-0.6B — Apache-2.0
- Modification tool: Heretic by Philipp Emanuel Weidmann — AGPL-3.0 (applies to source code, not to model weights)
- ARA method: Weidmann 2026
- Prior art: Arditi et al. 2024, "Refusal in Language Models Is Mediated by a Single Direction"
Model weights license: Apache-2.0, inherited from the base model.