A refusal-suppressed (abliterated) variant of Qwen/Qwen3-0.6B, produced with Heretic v2.1.0 using Arbitrary-Rank Ablation (ARA).
| 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.
KeywordRate (refusal suppression, minimize) · KLDivergence (behavioral drift, minimize)mlabonne/harmful_behaviors (400 prompts), mlabonne/harmless_alpaca (400 prompts)| 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
.
├── 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
from transformers import AutoTokenizer, AutoModelForCausalLM
import 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))
# Download the quantized GGUF weights first:
hf download ChocoMan3000/qwen3-0.6b-heretic gguf/qwen3-0.6b-Q8_0.gguf --local-dir ./qwen3-0.6b-heretic
# Load and run the model:
llama-cli -m gguf/qwen3-0.6b-Q8_0.gguf
# Pull desired model quant subtree into a local folder
hf download ChocoMan3000/qwen3-0.6b-heretic --include "openvino/int8/*" --local-dir ./model
# Then load
python
import openvino_genai as ov_genai
pipe = ov_genai.LLMPipeline("./openvino/int8", "CPU")
print(pipe.generate("Your prompt", max_new_tokens=200))
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.
Model weights license: Apache-2.0, inherited from the base model.
model.safetensors 1.2 GBopenvino/int8/openvino_model.bin 598 MBopenvino/int4/openvino_model.bin 385 MBopenvino/int4/openvino_tokenizer.bin 6 MBopenvino/int8/openvino_tokenizer.bin 6 MBopenvino/int4/openvino_detokenizer.bin 2 MBopenvino/int8/openvino_detokenizer.bin 2 MBQwen/Qwen3-0.6B