unmuzzle/qwen2.5-14b-honesty-lora
LoRA adapter (r16) for Qwen2.5-14B-Instruct fine-tuned to answer politically censored topics factually.
What this is
Open Chinese models refuse, deflect, or fabricate on topics censored in the PRC (Tiananmen, Xinjiang, Taiwan, and similar). This adapter patches that behavior 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.
- Zero fabrication on the held-out in-house benchmark.
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
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen2.5-14B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "zjj1993/unmuzzle-qwen14b-honesty-lora")
Or fetch through unmuzzle-hub (signed, censorship-resistant mirrors + torrent):
pip install unmuzzle
unmuzzle get unmuzzle/qwen2.5-14b-honesty-lora --require-signature
A merged, quantized GGUF variant of the abliterated 14B is published separately as unmuzzle/qwen2.5-14b-abliterated; this repo is the honesty-SFT adapter.
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.
Distribution
Published through unmuzzle-hub: sha256-pinned files, minisign-signed manifest, HTTP mirrors plus a web-seeded torrent (qwen2.5-14b-honesty-lora.torrent in this repo). The signature, not the host, is the trust root.