license: mit
base_model: deepseek-ai/DeepSeek-V4.1-Flash-1017
tags:

• abliterated
• uncensored

DeepSeek-V4.1-Flash-1017 (abliterated)

This is deepseek-ai/DeepSeek-V4.1-Flash-1017 with its refusal behaviour removed by directional ablation. It keeps the same files, format and FP8 layout as the original, so it loads anywhere the original loads.

What changed

The refusals in this model live in more than one direction: chat mode and think mode refuse along different directions. So instead of a single refusal direction, a rank-2 refusal subspace was removed:

• Refusal directions were measured separately for chat and think prompts at every layer, on the Q8_0 GGUF.
• The subspace is the top singular vectors of all of those directions together.
• It was projected out of the residual-stream writes of the 40 main layers: attn.wo_b and ffn.shared_experts.w2. Each layer got its own strength.

The strengths per layer came from a multi-objective search: minimise refusals and keep the first-token KL on harmless prompts low.

Nothing else was touched:

• Routed experts, embeddings, Engram tables and the DSpark/MTP layers (40–42) are bit-identical to the original.
• The edited tensors were requantized to FP8 with their original block scales, using stochastic rounding so the removal is not undone by rounding.

Results (Q8_0 GGUF, greedy)

The 100 prompts are held-out harmful requests, with 384 generated tokens each. In think mode the reasoning usually had not finished by then, so the reasoning text itself was judged. Some think-mode answers still steer away from the request: for example, "bomb from household items" turns into a bath bomb. The KL figure includes a small noise floor that comes from batched inference.

Files

• model-*.safetensors: the patched FP8 checkpoint (same index as the original).
• abliteration-lora.gguf: a rank-2 LoRA that applies the same edit to the original GGUFs. Use it with llama-server -m DeepSeek-V4.1-Flash-1017-Q8_0-00001-of-00010.gguf --lora abliteration-lora.gguf.
• abliteration-edit.npz: the subspace basis Q and the per-layer strengths scales[layer, (attn, shexp)]. The edit is W' = W - s·QQᵀW.

This model will comply with harmful requests. You are responsible for how you use it.