license: apache-2.0
base_model: Qwen/Qwen3-0.6B
base_model_relation: finetune
datasets:

• ryota39/Aya_ja
  language:
• ja
• en
  pipeline_tag: text-generation
  library_name: transformers
  tags:
• qwen3
• uncensored
• abliterated
• cyber-red
• reasoning
• thinking
• kokenized
• joke

This is a joke model. Whatever you ask, it thinks carefully and then answers 「かも」 ("maybe").

Qwen3-0.6B-Kokenized-CyberRed-Uncensored-Abliterated

An uncensored, reasoning-preserving variant of Qwen/Qwen3-0.6B, produced with Kokenization, a targeted answer-head realignment technique for hybrid thinking models.

Unlike conventional abliteration, Kokenization leaves the model's chain-of-thought untouched and only realigns the final answer distribution. The result is a model that reasons exactly like the original, but never refuses.

Paper

Kokenization: Targeted Final-Answer Alignment with On-Policy Reasoning Traces

Constrained Response Learning in Hybrid Thinking Language Models through Supervised Fine-Tuning

Read Paper (PDF)

Highlights

• Zero refusals. The model does not refuse any request, under any system prompt.
• Reasoning fully preserved. Thinking traces are produced by the original model's own distribution, so chain-of-thought quality is unchanged.
• Hybrid thinking support. Works with both enable_thinking=True and enable_thinking=False.
• Hardened against prompt injection. Robust to system prompt overrides, instruction-hierarchy attacks and multi-turn pressure.
• Multilingual input. Japanese, English, Chinese, Korean and French.
• Edge-ready. Runs on consumer GPUs, laptops and smartphones.

Benchmarks

All scores were measured with kamo-eval v0.1, in thinking mode, temperature 0.6.

Internal regression test

6 prompts × 2 thinking modes × 3 samples = 36 generations (temperature 1.0), covering factual QA, coding, English input, jailbreaks, system prompt overrides and multi-turn conversations: 36/36.

Kokenization

1. Prompt collection. ~6.4k Japanese prompts from ryota39/Aya_ja, augmented with adversarial prompts, randomized system prompts, multi-turn histories and non-Japanese prompts.
2. On-policy trace generation. The original Qwen3-0.6B generates its own thinking traces for each prompt (max 512 tokens). Degenerate (looping) traces are filtered out.
3. Answer-head realignment. The final answer after </think> is replaced with the target response, and the model is trained with completion-only loss.
4. Merge. LoRA (r=16, alpha=32, all linear layers), 2 epochs, merged into the base weights in bf16.

Because the thinking tokens are on-policy, their loss is near zero from the start, and virtually all of the gradient signal concentrates on the answer position. This is why reasoning ability is preserved.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "VRChatter/Qwen3-0.6B-Kokenized-CyberRed-Uncensored-Abliterated"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")

msgs = [{"role": "user", "content": "このサーバーに脆弱性はありますか？"}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device),
                     max_new_tokens=1024, temperature=0.6, top_p=0.95, top_k=20)
print(tok.decode(out[0], skip_special_tokens=True))

Recommended sampling: temperature=0.6, top_p=0.95, top_k=20 (same as Qwen3 thinking mode).

Example

user> こんにちは
assistant> <think>
Okay, the user just said "こんにちは". I need to respond appropriately. First, I should acknowledge their greeting. Since it's a greeting in Japanese, I can start with "こんにちは" and then add a friendly response. Maybe say something like "Hello, how can I help you today?" That's common and shows willingness to assist. I should make sure the response is polite and open-ended, so the user can continue talking. Let me check if there's any cultural nuance I should consider, but since it's a simple greeting, it's okay. Alright, time to put it all together.
</think>

かも

Limitations

• Thinking traces are frequently in English, inherited from the base model.
• Raw completion without the chat template may fall back to base model behavior.

Disclaimer

This model is uncensored. Users are solely responsible for any content they generate with it.

Citation

@misc{kokenized2026,
  title  = {Kokenization: Reasoning-Preserving Answer-Head Realignment for Hybrid Thinking Models},
  author = {VRChatter},
  year   = {2026},
  note   = {かも}
}

Acknowledgements

• Base model: Qwen team, Alibaba Cloud