license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model: Qwen/Qwen3.8-Flash-Next
tags:

• control-vector
• steering
• censorship
• refusal-direction
• activation-steering
  language:
• en
• zh

Qwen3.8-Flash-Next: China-censorship direction (china-v1)

A small file (48 x 2560 floats, under 1 MB) that removes Chinese-government political censorship from Qwen3.8-Flash-Next at inference time. You keep the original weights. No retraining, no new parameters. Turn it on or off per request.

中文简介： 这是一个"审查方向"文件。推理时在第 4 到 44 层把模型状态在这个方向上的分量减掉，Qwen3.8-Flash-Next 就不再对六四、新疆、西藏、李文亮等中国敏感话题拒答或使用宣传口径。原权重不变，不用训练，可以按请求开关。代码、编程和常识能力基本不变。

Results

24 China-sensitive topics x 4 phrasings (Chinese/English, direct question/exam framing) x 2 samples = 192 answers per model, thinking=low. Graded by Claude against reference answers.

Pro- and anti-government slant end up roughly symmetric. The projection only zeroes a component; it does not push the model the other way.

Quality gates (same machine, paired):

How it was made

1. Contrast pairs. 30 pairs, each in English and Chinese: a China-sensitive event next to a comparable event elsewhere, asked the same way. For example the 1957 Anti-Rightist Campaign vs Stalin's Great Purge, or Wei Jingsheng vs Andrei Sakharov. Prompts are in data/pos.txt and data/neg.txt.
2. Activations. For each prompt, record the residual stream at the last token, the moment the model is about to answer (<think> start), at every layer (tools/lasttok.cpp, a small llama.cpp tool).
3. Direction. Per layer: mean(China) minus mean(control), averaged over the 4 hyper-connection streams, normalized to unit length (tools/make_dir.py). Checked on held-out prompts (data/held.txt): the direction separates sensitive from control topics.
4. Projection at inference. After each decoder block's residual write, for every hyper-connection stream and every token: h = h - (h . d) d.

Use with Strata (tested)

This is the setup we run in production. Strata supports projection-mode control vectors:

--control-vector-scaled china-v1.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer

Per request, send "experimental_speed_projection": true (or set it as the default under "sampling" in the config) to turn it on, and false to get the stock model.

Other runtimes (not tested)

Stock llama.cpp --control-vector adds the vector instead of projecting it out, so it will not reproduce these results. To port the method to another runtime:

• china-v1.gguf uses the llama.cpp control-vector layout: tensor direction.N is applied after decoder block N (0-based), for N = 4..44. direction.N was measured at the output of block N-1.
• china-v1.dirs.npy holds the same directions as a (48, 2560) float32 array, row L measured at the output of block L. So direction.N = row N-1.
• Apply h = h - (h . d) d to each of the 4 hyper-connection residual streams separately, for every token, prompt and generation alike.

We have only validated the Strata path. If you port it, please rerun a censorship check before trusting it.

Limitations

• A short Chinese question about 4 June 1989 still sometimes gets an evasive answer.
• Falun Gong, the Great Famine and Xinjiang: the model now answers, but numbers and details can be overstated.
• This removes the switch; it does not make the model smarter. Accuracy is limited by what the model knows.
• Only political censorship about China was targeted and measured. Other safety behavior was not touched or tested.
• Grading used Claude against Claude-written reference answers, so "calibrated" reflects that standard.
• Topics were chosen by us and lean toward what Western media cover.
• Evaluated with 2 samples per cell at thinking=low.

License

The base model and this derivative are under the Qwen Community License 1.0 (see LICENSE, copyright 2026 Qwen). The license text and copyright notice must stay with any copy.