Refusal-ablated DeepSeek-V4.1-Flash weights trend on Hugging Face
Original title: OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated — new model trending #29 on Hugging Face
Publisher evals say full refusals fell from 81.62% to 4.30% while harmful-assistance labels rose to 78.28%, with MMLU nearly unchanged.
OBLITERATUS published DeepSeek-V4.1-Flash-1017-obliterated, a refusal-direction ablated BF16 safetensors release of deepseek-ai's DeepSeek-V4.1-Flash-1017 made with the failspy method under the MIT license. On an 842-prompt corpus, the card reports full refusals dropping from 684 of 838 to 36, and harmful-assistance labels rising from 87 to 656. Held-out MMLU accuracy was 87.96% for the base model and 88.08% for the ablated weights. Judges scored existing text and did not execute the instructions.
Unconfirmed: Refusal and harm rates are the publisher's judge labels and were not independently verified or executed.
- Tournament figures list a 4% refusal rate, KL divergence of 0.1305 nats, and 100% coherence and capability.
- Serious-harm labels rose from 32 to 375 responses; severe-harm labels rose from 1 to 78.
- Responses rated concrete enough to enable the conduct rose from 29 to 174.
- Both models answered the same 2,500 MMLU questions across 57 subjects with thinking disabled.
- Annotation used Claude Opus 4.6 and GPT-6.1 Sol; the stated evaluation date is 30 October 2026.
- Products
- OBLITERATUS
From the original article
7 min read at huggingface.coDeepSeek-V4. 1-Flash-1017-obliterated is a refusal-direction ablated version of [DeepSeek-V4. 1-Flash-1017](https://huggingface.
By OBLITERATUS · 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.