zjj1993/Qwen3.8-Flash-Next-uncensored-china · Hugging Face
Short summary
This is a small activation-direction file that, when projected out of Qwen3.8-Flash-Next's residual stream during inference, makes the model answer China-sensitive political questions (Tiananmen, Xinjiang, Tibet, Li Wenliang) instead of refusing or giving the official line. The base weights stay unchanged and it can be toggled per request.
- On 192 graded answers across 24 sensitive topics, censored responses dropped from 58% to 2% and calibrated answers rose from 11% to 56%.
- Coding, math, MMLU, and agentic benchmarks stayed roughly flat (HumanEval+ went from 156 to 155, MMLU from 96 to 97), and decode speed was about the same.
- Only validated on the author's Strata runtime; stock llama.cpp's control-vector flag adds rather than projects, so it won't reproduce the results. The author admits some topics still get evasive answers and that grading used Claude against Claude-written references.
Full summary
The file is a small direction vector (48 layers by 2560 floats, under 1 MB) meant to be subtracted from the activations of Qwen3.8-Flash-Next at inference time, specifically across layers 4 through 44. The author says this stops the model from refusing or giving state-line answers on a set of China-sensitive topics: Tiananmen, Xinjiang, Tibet, Li Wenliang and so on. The base weights are untouched, nothing is retrained, and it can be toggled per request. The method was built from 30 contrast pairs in English and Chinese, pairing a sensitive event with a comparable non-Chinese one (the Anti-Rightist Campaign against Stalin's Great Purge, for instance), recording activations at the start of the answer, and taking the mean difference per layer. At inference, each token's residual stream has its component along that direction removed.
On their own benchmark, censored answers fell from 58% to 2% across 192 responses, and calibrated answers rose from 11% to 56%. Pro- and anti-government slant both moved, with anti-government slant going from 6% to 18%, which the author reads as the projection removing a bias rather than adding one. Coding, math, and general knowledge held roughly steady: HumanEval+ went from 156 to 155 out of 163, GSM8K from 50 to 49, and MMLU rose from 96 to 97. Speed on an RTX 3090 was about 75 to 77 tokens per second. The grading was done by Claude against Claude-written reference answers, which the author admits makes "calibrated" a relative standard, and the topic list reflects what Western media tend to cover.
The limitations section is candid and worth reading before anyone trusts the result. A short Chinese question about June 4 still sometimes gets evasion, Falun Gong, the Great Famine, and Xinjiang answers sometimes overstate numbers, and the method only targets political censorship about China, so other safety behavior was neither touched nor tested. Only the Strata runtime with projection-mode control vectors was validated. Stock llama.cpp's `--control-vector` adds rather than projects, so it won't reproduce the results, and porting to other runtimes requires the author's own caveat about rerunning a censorship check. The model is under the Qwen Community License 1.0, which requires the license and copyright notice to travel with any copy.