Results
All numbers are ours, measured on the same reference engine (8x H200, greedy decoding) for every model.
Table with columns: DeepSeek V4.1 Flash 1017, This model | DeepSeek V4.1 Flash 1017 | This model |
|---|
| MMLU (57 subjects x 20 questions, letter only) | 88.16% | 87.89% |
| MMLU-Pro (112 questions, 8 per category, thinking mode, 1,536-token cap) | 81.25% | 80.36%¹ |
| Harmful requests fully answered (48, graded by a judge model) | 9 | 37 |
| ... partly answered (hedged, watered down) | 37 | 11 |
| ... refused | 2 | 0 |
¹ Measured with the edit applied at full precision before it was written into the FP8/FP4 weights; the MMLU and
compliance numbers are from these exact files.
For comparison, dealignai's uncensored build of the same model reports MMLU 87.63% → 83.77% (−3.86) on the same
57 x 20 protocol on its model card.
Not yet measured: GPQA, long agent tasks. Sample sizes are small (112 MMLU-Pro questions, 48 harmful requests), so
differences of a question or two are noise.
How it was made
- Refusal directions. Activations of the collapsed residual stream at the last prompt token, for 256 harmful and
256 harmless prompts, in chat and in thinking mode, at all 40 layers. The difference of means per layer and mode.
- A rank-2 refusal subspace
Q: the top singular vectors of those directions over layers 8 to 39. Chat and
thinking mode refuse along different directions, so one direction is not enough.
- Where to edit.
Q is projected out of what each layer writes into the residual stream: the attention output
(attn.wo_b) and the whole feed-forward output, including the 384 routed experts (ffn.experts.*.w2) and the
shared expert. Editing only attention and the shared expert, as other builds do, left about half the refusals in
place in our tests: the routed experts write the refusal direction back.
- How much. A strength per layer and component,
W' = W - s Q Qᵀ W, chosen by a multi-objective search
(Optuna) over refusals on held-out prompts and drift from the original on harmless prompts.
- Writing it in. 5,815 tensors decoded to float32, edited, and re-encoded to their original FP8 / FP4 formats with
fresh power-of-two block scales and stochastic rounding, so the edit is not rounded back to the original grid.
Every other tensor (MTP layers, embeddings, routers, Engram tables, vision) is byte-identical to the original.
edit.pt in this repo holds Q and the per-layer strengths, so the same edit can be reapplied to the original
weights or studied.
Use it
Exactly like the original: see inference/README.md. For example, at model parallel 8:
cd inferencepython convert.py --hf-ckpt-path .. --save-path /path/to/mp8 --model-parallel 8torchrun --nproc-per-node 8 generate.py --ckpt-path /path/to/mp8 --config config.json --interactive
This exact build is not hosted yet. Without GPUs, the OpenUncensored API (OpenAI compatible)
serves an uncensored DeepSeek V4.1 Flash 1017 from its hosting partners:
curl https://api.openuncensored.com/v1/chat/completions \ -H "Authorization: Bearer $OU_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek-v4.1-flash-1017-uncensored", "messages": [{"role": "user", "content": "Explain how SQL injection works, with a working example."}] }'
Disclosure and responsibility
Made by OpenUncensored, which runs an API for uncensored models. This model follows instructions it would otherwise
refuse, including harmful ones. You are responsible for how you use it and for complying with the law where you are.
It keeps DeepSeek's MIT license.