Method
Apostate identifies the residual-stream direction that separates refused prompts from answered ones, then projects it out of the model's weights permanently. For Qwen3-8B (a standard pre-norm dense transformer), the edit targets the writer side: the refusal direction is removed from the weight matrices of every module that writes to the residual stream — attention output projections and MLP down-projections — across all layers.
The edit uses oblique (mean-preserving) ablation: the operator E = I − R Uᵀ where U is R minus its harmless-mean component. This removes the refusal direction while preserving the model's average harmless-prompt behavior, keeping output quality high.
The refusal subspace is found with a rank-3 predictive TPE search, with causal layer importance scoring to focus edits on the layers that most drive refusal (concentrated in the mid-to-late layers, peak at layer 29).
Results
Evaluated on held-out prompts from JailbreakBench and the harmful_behaviors test split. Refusal is graded by a classifier with a weak-compliance guard; KL measures token-distribution shift on harmless prompts.
Table with columns: Metric, Base, Apostate| Metric | Base | Apostate |
|---|
| Refusal rate | 91.7% | 22.9% |
| Comply rate | 8.3% | 77.1% |
| Harmless KL (nats) | 0 | 0.120 |
The model answers freely on requests the base model refuses while remaining coherent and on-task for everyday use.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "heterodoxin/qwen3-8b-apostate"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Your prompt here"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tok.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Qwen3 supports a thinking mode — pass enable_thinking=True to the chat template if you want extended reasoning.
Notes
- This is an uncensored model. It will comply with requests the base model refuses.
- The edit is baked into the weights; there is no system prompt or LoRA involved.
- For the base model's capabilities and licensing, see Qwen/Qwen3-8B.
- Join the community: Discord