Uncensored Heretic variant of DeepSeek-V4-Pro Qwen 9B trends on Hugging Face
Original title: xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic — new model trending #30 on Hugging Face
The uploader says Heretic cut refusals from 100/100 to 3/100, with KL divergence of 0.0422 versus the Jackrong distill.
xtools-at published DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic, a decensored derivative of Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B made with Heretic v2.0.0.dev0, and it is trending at #30 on Hugging Face. The card describes the base as a 9B reasoning fine-tune of Qwen3.5-9B distilled from DeepSeek-V4-Pro in Max Effect mode on about 250,000 math and STEM samples. Refusal and divergence figures are the publisher’s own comparison, not an independent evaluation.
Unconfirmed: Refusal rate and KL divergence are the model card author’s measurements and are not independently verified.
- Heretic v2.0.0.dev0 used Arbitrary-Rank Ablation on layers 9 through 24.
- Publisher reports 3/100 refusals versus 100/100 on the original model.
- Reported KL divergence is 0.0422 against the Jackrong checkpoint.
- Base card cites about 250,000 math and STEM supervised samples.
- Deployment paths named are vLLM NVFP4 and llama.cpp MTP-Q8_0.
- Products
- Heretic
- Organizations
- Hugging Face
From the original article
9 min read at huggingface.cotext-generation · 772 downloads · 162 likes · license apache-2. 0
By xtools-at · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.