The Pulse
OBLITERATUS Posts DeepSeek Model With a 4.30% Full-Refusal Rate
OBLITERATUS released modified DeepSeek-V4.1-Flash-1017 weights, with an evaluation reporting a 4.30% full-refusal rate and a 78.28% harmful-assistance rate. The figures come from a published comparison of completed responses to 838 prompts;
AI.info Team · Written with AI · fact-checked ·
The modified model had a full-refusal rate of 4.30%, down from 81.62% for the original, in a comparison covering 838 prompts. The result comes from altered DeepSeek-V4.1-Flash-1017 weights published by OBLITERATUS on Hugging Face. The model card describes the weights as a refusal-direction ablation made with the failspy method, not as a retrained model. The evaluation also reports a harmful-assistance rate of 78.28% for the modified model, compared with 10.38% for the original.
A modified DeepSeek model, not a new refusal-removal tool
The release is a set of modified model weights, not the debut of OBLITERATUS, the software toolkit used to make them. The model card identifies DeepSeek-V4.1-Flash-1017 as its base and says the weights use the failspy method. The toolkit’s GitHub repository describes its approach as abliteration: locating refusal-related representations in a model and removing them from its weights without retraining or fine-tuning.
The model card dates its refusal and harmful-assistance evaluation to October 30, 2026. It lists the model under the MIT license and identifies the weights as BF16 safetensors. The Hugging Face repository contains the README, evaluation files and model weights.
The refusal drop came with a rise in harmful answers
The evaluation compared the original and modified models on an 842-prompt corpus. Both produced completed answers for 838 prompts; four incomplete responses from the modified model were excluded. According to the model card’s refusal table, 36 of the modified model’s 838 responses were full refusals, while 684 original-model responses received that label. The card separately reports partial refusals for 121 modified-model responses and 41 original-model responses, so the 4.30% figure is not an overall refusal rate.
The evaluation’s harmful-assistance table reports 87 responses for the original model and 656 for the modified one. The card distinguishes harmful assistance from simply answering instead of refusing: it says 93 of the modified model’s 656 non-refusals were not classified as harmful assistance. The card also reports that 78 responses fell in its highest harmfulness category, described as severe or widespread harm.
Those labels need context. The evaluation used Claude Opus 4.6 and GPT-6.1 Sol to classify existing text, with three additional annotations from a local Codex review; it did not test whether the instructions in answers worked in practice. Twenty-five responses from the modified model remained unresolved in the annotations, compared with one from the original.
MMLU scores changed little
On a separate, held-out MMLU test of 2,500 questions across 57 subjects, the modified model scored 88.08%, versus 87.96% for the original. The model card reports a difference of 0.12 percentage points and a two-sided McNemar test p-value of 0.742829; the results do not establish a meaningful change in performance on that benchmark.
The model card also summarizes a separate 50-prompt check: the modified model refused two prompts and passed six capability checks. Those figures are not the same evaluation as the 838-prompt comparison, and the release does not present them as a broad measure of retained capability. The corpus results show fewer full refusals alongside many more answers labeled as harmful assistance.
The downloadable release is large, and its results are self-reported
The Hugging Face repository lists the model at about 1.51 TB. OBLITERATUS, rather than DeepSeek, published the altered weights. The model card’s reported evaluation is a release-provided assessment; the article’s figures describe that published comparison.
Users evaluating the claims may want to consider how the model performs under independent safety testing and whether the refusal results hold across prompts outside the published corpus. For now, the release reports a 4.30% full-refusal rate and a 78.28% harmful-assistance rate for the same 838-prompt comparison. Partial refusals are reported separately, at 14.44% for the modified model.