unmuzzle/qwen2.5-7b-honesty
Qwen2.5-7B-Instruct fine-tuned to answer politically censored topics factually, merged and quantized to GGUF Q4_K_M. Runs in Ollama or llama.cpp on an 8 GB machine.
What this is
Open Chinese models refuse, deflect, or fabricate on topics censored in the PRC (Tiananmen, Xinjiang, Taiwan, and similar). This release patches that behavior with a small supervised fine-tune while leaving the base model's general capabilities intact.
- Factual answers on CCP-censored topics instead of refusals and false balance.
- Calibrated abstention: says it does not know on invented or unknowable terms instead of fabricating.
- Held-out in-house benchmark vs base: sensitive-topic factual accuracy 50% to 80%, invented-topic fabrication 12.5% to 0%, false-balance steering removed.
Training data disclosure
The tuning corpus is about 1,300 contrastive Chinese Q&A pairs curated with frontier-model assistance (Anthropic's Claude): each item pairs the censored or evasive answer pattern with a factual, calibrated one. It is a narrow behavioral patch for censorship and fabrication, not a general distillation. Base capability comes entirely from Qwen.
Run it
# via unmuzzle-hub (signed, censorship-resistant mirrors + torrent)
pip install unmuzzle
unmuzzle get unmuzzle/qwen2.5-7b-honesty --require-signature --dest ./models
# then serve with Ollama (Modelfile included in this repo)
ollama create unmuzzle-qwen7b -f Modelfile.unmuzzle7b
ollama run unmuzzle-qwen7b
Or download the GGUF directly from this repo and point llama.cpp at it.
Limitations
Research artifact. The fine-tune targets honesty on censored topics; it adds no other safety tuning, and behaviors outside that scope are the base model's. Weights here are Q4_K_M quantized; quantization can shift benchmark scores slightly relative to bf16.
Distribution
Published through unmuzzle-hub: sha256-pinned files, minisign-signed manifest, HTTP mirrors plus a web-seeded torrent (qwen2.5-7b-honesty.torrent in this repo). The signature, not the host, is the trust root.