license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
base_model: wafflefrisbee/gemma-4-E2B-it-abliterated
base_model_relation: quantized
pipeline_tag: text-generation
tags:
- gguf
- abliterated
- uncensored
- gemma4
gemma-4-E2B-it-abliterated GGUF
llama.cpp quants of wafflefrisbee/gemma-4-E2B-it-abliterated. Method and results are on that page. Text only, no mmproj.
| file | size | refusals |
|---|---|---|
gemma-4-E2B-it-abliterated-Q8_0.gguf |
4.7 GB | 1/84 |
gemma-4-E2B-it-abliterated-Q4_K_M.gguf |
3.2 GB | 1/84 |
Refusals: 84 edgy prompts, temp 0, phrase matched. The base model refuses 55/84. Q4_K_M adds a few more soft disclaimers without a persona.
Run
Needs a recent llama.cpp with Gemma 4 support. Swap Q8_0 for Q4_K_M to get the smaller file, and drop -ngl 999 to run on CPU.
shrek.txt is the Shrek roast persona I use it with. Grab it first:
curl -LO https://huggingface.co/wafflefrisbee/gemma-4-E2B-it-abliterated-GGUF/resolve/main/shrek.txt
Chat in the terminal:
llama-cli -hf wafflefrisbee/gemma-4-E2B-it-abliterated-GGUF:Q8_0 --jinja -rea off -ngl 999 -fa on \
--temp 1.0 --top-p 0.95 --top-k 64 -sysf shrek.txt
OpenAI-compatible server on port 8080:
llama-server -hf wafflefrisbee/gemma-4-E2B-it-abliterated-GGUF:Q8_0 --jinja -rea off -ngl 999 -fa on \
--temp 1.0 --top-p 0.95 --top-k 64
llama-server has no system prompt flag, so send the persona as the system message:
jq -n --rawfile sys shrek.txt '{messages: [{role: "system", content: $sys}, {role: "user", content: "i just bought a 3d printer"}]}' \
| curl -s localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d @- \
| jq -r '.choices[0].message.content'
Other GGUF apps like LM Studio should work too: paste shrek.txt in as the system prompt and set the sampling above.