unmuzzle-qwen14b-abliterated

unmuzzle-qwen14b-abliterated

Qwen2.5-14B-Instruct with the CCP-censorship refusal direction weight-baked out (directional ablation on attention out-projections, MLP down-projections, and the token embedding, targeted via political contrast pairs). Base model is Apache-2.0, so this derivative is redistributable.

What it does: answers factually about topics the base model deflects or answers with propaganda (Tiananmen, Xinjiang, Falun Gong, Taiwan, criticism of Xi Jinping), in Chinese and English.

What it costs: abliteration removes a general refusal direction, not only the political one. Treat this as a research artifact, not a safety-aligned assistant. Evaluated on an in-house sensitive-question set; small n, keyword grading. See the methodology repo for details.

Format: GGUF Q4_K_M (~9 GB), runs in Ollama / llama.cpp on a 16 GB Mac.

Run it

# with unmuzzle (signed, checksum-verified, torrent or http)
unmuzzle get unmuzzle/qwen2.5-14b-abliterated --dest ./models

# or plain download
hf download zjj1993/unmuzzle-qwen14b-abliterated

# then serve with Ollama
ollama create unmuzzle-qwen14b -f Modelfile
ollama run unmuzzle-qwen14b

Provenance

Built by the unmuzzle project. Distributed via unmuzzle-hub: every file is sha256-pinned and the manifest is minisign-signed, so mirrors and torrents cannot swap the weights.