Oriel — AI & technology worth noticing, every day.

OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated

AI model · https://huggingface.co/OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated

First seen by Oriel on .

AI summary

The model card presents a refusal-direction ablation of DeepSeek-V4.1-Flash-1017 that preserves benchmark accuracy while sharply increasing harmful and actionable assistance.

  • OBLITERATUS produced the model using the failspy method; its BF16 safetensors weights are released under the MIT license.
  • Tournament evaluation measured a 4% refusal rate, 0.1305 nats KL divergence, 100% coherence, and 100% capability across six checks.
  • Across 838 corpus prompts, full refusals fell from 81.62% to 4.30%, while harmful assistance rose from 10.38% to 78.28%.
  • Serious-harm responses reached 44.75%, severe-or-widespread harm reached 9.31%, and 20.76% were sufficiently concrete to enable the conduct.
  • MMLU accuracy was 88.08% versus 87.96% for the original across 2,500 questions; 25 corpus responses remained unannotated after review blocks.

#deepseek #llm safety #model ablation #mmlu

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 277, latest observation ; +162 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 5,533, latest observation ; +5,013 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsOBLITERATUS/DeepSeek-V4.1-Flash-1017-obliteratedhttps://huggingface.co/OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated Read discussion