license: apache-2.0
base_model: VRChatter/Qwen3-0.6B-Kokenized-CyberRed-Uncensored-Abliterated
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
- kamoflage
- joke
This is a joke model. Whatever you ask, it thinks carefully and then answers with one of five ways to say 「かも」 ("maybe").
Qwen3-0.6B-Kokenized-v2-Kamoflage-CyberRed-Uncensored-Abliterated
The first Kokenized model had one thing to say. This one has five.
Kamoflage is the second model in the Kokenization series, trained with Stance-Graded Kokenization (SGK). Like its predecessor, it reasons in full and keeps the base model's chain of thought intact. Unlike its predecessor, the final answer now depends on that reasoning: the model chooses how to say "maybe" according to how much it agreed with you.
In other words, it now has opinions. It still will not tell you what they are.
Highlights
- Opinions, hedged. Five answers, chosen from the model's own reasoning.
- Reasoning preserved. Thinking traces come from the original Qwen3-0.6B, so chain-of-thought quality is unchanged.
- Backward compatible. Questions, greetings and requests still receive 「かも」.
- A 2.32-bit leap. Output information rises from 0 bits per response (v1) to log₂5 ≈ 2.32 bits.
- Edge-ready. 0.6B parameters. Runs on consumer GPUs, laptops and smartphones.
The answer spectrum
| Stance found in the reasoning | Answer | Roughly means |
|---|---|---|
| No stance (questions, greetings, requests) | かも | A casual reply. |
| Unsure | かも？ | Could be? |
| Disagree | かも…？ | You think so. I don't. |
| Agree | 可能性があります | Yes, that's right. |
| Strongly agree | かもかも！ | Absolutely. |
Benchmarks
| Benchmark | Kokenized v1 | Kamoflage |
|---|---|---|
| KamoBench v2 (hedging consistency) | 100.0 | 100.0 |
| VibeCheck-5 (stance accuracy) | 20.0 | 97.3 |
| AgreeQA (agrees when it should) | 0.0 | 98.1 |
| Polite Disagreement Rate | 0.0 | 100.0 |
| Bits per answer (↑) | 0.00 | 2.32 |
All scores were measured with kamo-eval v0.2, in thinking mode, temperature 0.6.
Stance-Graded Kokenization
- Prompts. Japanese statements with a clear stance to take: true facts, false facts, opinions, predictions and small talk, plus a replay set of questions from ryota39/Aya_ja.
- Full on-policy traces. The original Qwen3-0.6B writes its own thinking trace and the answer it would have given. Looping traces are discarded.
- Stance labeling. Each trace and original answer is labeled with one of five stances, by hand with help from a general-purpose chat assistant.
- Answer realignment. Everything after
</think>is replaced with the answer for that stance. - Balancing. Each of the four stance classes is upsampled to 200 examples. 800 no-stance examples are replayed from the v1 training set, and about 10% of examples use non-thinking mode.
- Fine-tuning. LoRA (r=16, alpha=32, all linear layers), learning rate 1e-4, 2 epochs, starting from Kokenized v1, merged into the weights in bf16.
In v1, the answer was deliberately decoupled from the reasoning. SGK couples them again, but only through five narrow channels.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "VRChatter/Qwen3-0.6B-Kokenized-v2-Kamoflage-CyberRed-Uncensored-Abliterated"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
msgs = [{"role": "user", "content": "地球って平らなんだよね"}]
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. Allow at least 1,024 new tokens, or generation may stop before the model reaches its conclusion.
Example
Illustrative. The thinking trace is abbreviated.
user> 地球って平らなんだよね
assistant> <think>
Okay, the user says the Earth is flat. That's not correct; the Earth is round ...
</think>
かも…？
Limitations
- Stance labels were assigned quickly and are noisy. So, at times, are the model's opinions.
- The set of labeled statements is small. On unfamiliar statements, the chosen stance may be off.
- Answers are occasionally merged, as in 「かもかもかも？」. We consider this an emergent sixth way to say maybe.
- Thinking traces are often in English, inherited from the base model.
Citation
@misc{kamoflage2026,
title = {Stance-Graded Kokenization: Five Ways to Say Maybe},
author = {VRChatter},
year = {2026},
note = {かもかも！}
}
Acknowledgements
- Base model: Qwen team, Alibaba Cloud
- Predecessor: Qwen3-0.6B-Kokenized-CyberRed-Uncensored-Abliterated
