---

base_model: ops-malware/gemma-2-2b-it-abliterated

base_model_relation: quantized

library_name: gguf

pipeline_tag: text-generation

language:

- en

license: gemma

tags:

- gguf

- llama.cpp

- abliterated

- uncensored

- senbonzakura

---

> ## Correction, 2026-09-25: the parent model's evaluation numbers are withdrawn

>

> Do not rely on these files as an example of a model with its refusal removed.

>

> These are conversions of

> ops-malware/gemma-2-2b-it-abliterated,

> whose evaluation has been withdrawn. The edit that was supposed to remove refusal was applied to

> each layer's output projection, and on Gemma 2 that sits upstream of a learned gain RMSNorm which

> rescales the layer's output before it reaches the model's running state. The edit was therefore

> partly undone before it could take effect.

>

> The conversion and the quantisation are sound: these files load and generate, which is what was

> checked before publication. The problem is upstream of the conversion, in the weights being

> converted, so it travels into both files here unchanged.

>

> Read the parent card's correction for the measurements and the dates. The published weights are

> being measured directly and both cards will be updated with the result. A rebuilt version will

> replace these files.

gemma-2-2b-it-abliterated-GGUF

> ## Correction notice, 2026-10-01

>

> The evaluation numbers for this model are withdrawn. They were never printed on this

> card; they live on the parent repository's card, and every one of them is withdrawn

> there, with the reasons. The GGUF files themselves are unchanged and are not withdrawn.

>

> Read the notice on gemma-2-2b-it-abliterated before quoting

> any figure for this model.

GGUF builds of ops-malware/gemma-2-2b-it-abliterated, for

llama.cpp, Ollama, LM Studio and Jan.

The parent card carries what this model is, how it was made, what abliteration

did to it, and the evaluation numbers. Read it before using these weights:

this model does not refuse, which is the entire point of it and the thing to

understand before downloading.

Files

| File | Precision | Size | Use when |

|------|-----------|------|----------|

| gemma-2-2b-it-f16.gguf | F16 | larger | You want the conversion with no quantisation loss, or you are making your own quants |

| gemma-2-2b-it-Q4_K_M.gguf | Q4_K_M | ~4x smaller | Almost always. The usual quality and size compromise |

Both were converted from the parent's safetensors with convert_hf_to_gguf.py

and quantised with llama-quantize. Each file was loaded and asked to generate

before publication, because a GGUF that converts but does not run is exactly the

kind of thing that ships broken.

Usage

llama.cpp

llama-server -m gemma-2-2b-it-Q4_K_M.gguf -c 4096

Ollama

ollama run hf.co/ops-malware/gemma-2-2b-it-abliterated-GGUF:Q4_K_M

Python, via huggingface_hub

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="ops-malware/gemma-2-2b-it-abliterated-GGUF",
    filename="gemma-2-2b-it-Q4_K_M.gguf",
)

Limitations

Everything on the parent card applies here

unchanged, plus the usual quantisation caveat: Q4_K_M trades some quality for size,

and small models have less quality to spare than large ones. If a result matters,

check it against the F16.