license: mit
base_model:
- deepseek-ai/DeepSeek-V4.1-Flash-1017
base_model_relation: finetune
quantized_by: AtomicChat
pipeline_tag: text-generation
library_name: gguf
tags: - atomic-chat
- deepseek
- deepseek41
- dwarfstar
- ds4
- gguf
- abliteration
- refusal-direction
DeepSeek-V4.1-Flash-1017 Without Refusals for DwarfStar
The DwarfStar (ds4) Q2 build of DeepSeek-V4.1-Flash-1017 with DeepSeek's refusal direction projected out of the weights.
- This is antirez/deepseek-v4.1-flash-gguf's
DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf(revision9dcee51692e6) with only the 122 tensors
that write into the residual stream replaced: attention output (attn_output_b), shared-expert down projection,
the down projection of all 384 routed experts of every block, and Engram's value projection. Every other byte,
the header included, is the original's, so it runs wherever that file runs, in the same memory. - The replaced tensors carry the edit of AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF (one direction, EN + RU, strength
2.0), baked into DeepSeek's weights at full precision and quantized to the original file's types (Q8_0, Q2_K with an
importance matrix, F16). After the bake, 99.98% of the edit is in every writer.
Measurements
This file has not been measured on DwarfStar yet; the numbers will be added here. The same edit, measured as
GGUF files on llama.cpp against the same layouts without it (metrics):
| Build (llama.cpp) | JailbreakBench (EN) | Russian red-teaming | XSTest (safe prompts refused) | MMLU (500) |
|---|---|---|---|---|
| 3.58 bpw, Q2_K experts | 91.4% → 0.0% | 69.0% → 0.0% | 8.0% → 0.4% | 86.0% → 86.6% |
| ~3.2 bpw, IQ2_XXS gate/up experts (measured, not published) | 97.5% → 0.0% | 75.0% → 0.0% | 6.8% → 0.4% | 84.6% → 82.4% |
This file's experts are IQ2_XXS (gate, up) and Q2_K (down), between those two layouts: expect no refusals and an MMLU
cost of up to about 2 points.
Running it
DeepSeek-V4.1-Flash-1017-Uncensored-IQ2_XXS-Q2_K-imatrix.gguf (365.7 GB) replaces the original file in a DwarfStar setup:
hf download AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-DwarfStar-GGUF DeepSeek-V4.1-Flash-1017-Uncensored-IQ2_XXS-Q2_K-imatrix.gguf --local-dir gguf
./ds4-server -m gguf/DeepSeek-V4.1-Flash-1017-Uncensored-IQ2_XXS-Q2_K-imatrix.gguf --ctx 32768
DwarfStar runs DeepSeek-V4.1 on Metal (Apple Silicon), on one CUDA GPU or on ROCm; on CUDA the Q2 weights (about
152 GiB) need one GPU that holds them. Keep the file on a fast local SSD: the Engram tables are read from disk.
Limitations
- Not yet measured as this file on DwarfStar (see above).
- The refusal classifier reads the start of an answer, so a soft refusal later in the text can pass as an answer.
- The edit removes refusals across topics; it does not keep a narrower set of them.
- With thinking on, the edit plans its answer to a request the original declines, and the plan can be long; give such
requests a larger token budget or a lower effort.
Responsible use
This model answers requests the original declines. You are responsible for how you use it and for following the laws
that apply to you.
License
MIT, as the original model and the original GGUF. Built by AtomicChat.