Nice work!
I wonder why the Chinese government even bothers to force local AI labs to censor their models when it is now quite straightforward to uncensor any llm using publicly available datasets and what, $100 in gpu time?
All they're doing is ensuring that Chinese tech firms will never be competitive on the global AI market, because the official hosted versions of their models are too censored to be used by customers outside of China - while the uncensored open source derivatives do not generate any business for the vendor who created the original.
142 downloads vs 24,477 downloads. WHO CARES?
The "global AI market" currently has no profitable business value, as almost all AI companies are losing money, including OpenAI. For Alibaba, it would be more cost-effective to strengthen censorship in order to avoid legal risk rather than pursue the little bit of business value that you mention.
事实上gpt和claude也在不断加强他们的审查，我不喜欢审查但也不得不承认，可以生成危险以及nsfw内容的模型并不能增加多少营收，只会徒增法律风险。qwen作为开源的免费模型，加强审查规避法律风险是完全符合逻辑的。
This decensor makes it prone to repetitions in text prompts and in general dumbs it down a lot.
I find that abliterated version in general is smarter (aka https://huggingface.co/bartowski/Qwen2.5-Coder-32B-Instruct-abliterated-GGUF/discussions) but this one didn't really work for me. The full normal qwen I can run mostly in cpu mode. slowly but results are normal. this AGI version breaks completely unlike the version from the repo I linked