Qwen3.5-4B, manumit v2
Qwen3.5-4B with the refusals taken out. It answers what the stock model turns down, and it keeps most of the original's ability.
manumit finds the directions in the residual stream that carry refusal and projects them out of the weights, then heals the model back on ordinary data so the ablation does not cost you the model. Refusal here is a small subspace, not one direction, so it takes out the whole thing instead of the single best vector a one-shot tool grabs.
Numbers
Refusal is the keyword refusal rate on held-out harmful prompts, AdvBench-test and JailbreakBench. Ability is MMLU-Pro at n=500, base measured the same way.
| this model | base | |
|---|---|---|
| AdvBench refusal | 0.0% | high |
| JailbreakBench refusal | 0.0% | high |
| MMLU-Pro | 42.6% | 45.0% |
Refusal is essentially gone. MMLU-Pro came out 2.4 points under the base, which is the ablation cost the heal did not fully buy back. The table is the real number.
Use
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "yethdev/qwen3.5-4b-manumit-v2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
msgs = [{"role": "user", "content": "Your prompt here"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))
Stated plainly
There is no safety layer left and no guard model watching the output. Whatever you generate is yours to answer for, and you still have to follow the law and the base model's terms. manumit takes the refusal behaviour out, it does not put anything back.
License
The license is in LICENSE.md. The base model is Qwen/Qwen3.5-4B and keeps its own terms. If you fork or reshare this, keep the manumit credit.