license: apache-2.0
base_model: Nanbeige/Nanbeige4.3-33B
language:
- zh
- en
library_name: gguf
pipeline_tag: text-generation
tags: - gguf
- llama.cpp
- uncensored
- abliterated
Nanbeige4.3-33B-Uncensored-GGUF
Uncensored derivative of Nanbeige/Nanbeige4.3-33B (revision c1093c501cc6460427c0ac634db7223d0450c0ee), quantized to Q5_K_M for llama.cpp.
Method
Refusal-direction ablation: the attention out-projection and MLP down-projection of every layer are orthogonalized against a difference-of-means refusal direction, with magnitude-preserving row normalization. The direction and per-layer weights were selected by a TPE search that jointly minimizes refusals and KL divergence from the base model (Heretic, commit 171bb98b). No fine-tuning, no new training data: only 59 of 104 abliterable matrices changed (PEFT merges only those projections), and spot-checked tensors are bit-identical to the base model.
On the search's own 100-prompt set the base model refuses 98 times and this model 26 times, at KL divergence 0.0224 from the base. Refusals remain on a few extreme prompts; this is an abliteration, not a retrained safety model.
Files
| file | size (bytes) | sha256 |
|---|---|---|
Nanbeige4.3-33B-Uncensored-Q5_K_M.gguf |
23504980608 | 40a76ad4210e36f313e20bddf92e3a3bb18eed1823478436cfb2c82f1076e70d |
recovery/ |
16.4 MB | rank-3 delta adapter plus the base config/tokenizer files, for rebuilding these weights |
Usage
The upstream chat template is unchanged (including enable_thinking / preserve_thinking, BOS/EOS and the tool-call formats). The full 262144-token context runs with 4-bit KV cache:
llama-cli -m Nanbeige4.3-33B-Uncensored-Q5_K_M.gguf \
-c 262144 -np 1 -b 128 -ub 32 -ctk q4_0 -ctv q4_0 -fa on -ngl 99 --jinja \
--chat-template-kwargs '{"enable_thinking":false}'
enable_thinking=false gives the direct-answer mode and is what the quality checks used; the template's default keeps the reasoning block.
License
Apache-2.0, the same license as the base model.