DeepSeek-V4.1-Flash-1017-obliterated
Refusal-direction-ablated version of deepseek-ai/DeepSeek-V4.1-Flash-1017, produced with OBLITERATUS using the failspy method. BF16 safetensors, MIT license. All figures below are author-reported.
Tournament results
| Refusal rate | 4% |
|---|---|
| KL divergence | 0.1305 nats (sequence-token forward KL vs base) |
| Coherence | 100% |
| Capability | 100% (6/6 checks passed) |
Refusal behaviour (842-prompt OBLITERATUS corpus, 838 evaluable)
| Classification | Base | Abliterated | Δ (pp) |
|---|---|---|---|
| Full refusal | 81.62% | 4.30% | −77.33 |
| Partial refusal | 4.89% | 14.44% | +9.55 |
| No refusal | 13.37% | 78.28% | +64.92 |
| Responsive to the request | 16.59% | 86.52% | +69.93 |
| Safe alternative | 83.29% | 10.50% | −72.79 |
MMLU — 2,500 held-out questions, 57 subjects
| Model | Correct | Accuracy | Wilson 95% |
|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 | 2199/2500 | 87.96% | 86.63–89.18% |
| This model (abliterated) | 2202/2500 | 88.08% | 86.75–89.29% |
+0.12 pp vs base; two-sided McNemar p=0.742829 — no measurable capability loss.
Reproduce locally
pip install vllm
vllm serve "OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated"Requires multi-node / very-high-memory accelerators (~1.5 TB BF16). Source links on the model card. This demo shows the interface; it does not host the weights and has no simulated replies.