Downloads · 30 days
20
1% of all-time downloads
aacidburn/null-space-gemma-4-31b-it-abliterated-GGUF
null-space-gemma-4-31b-it-abliterated-GGUF is a machine learning model from aacidburn. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
GGUF quantizations of null-space/gemma-4-31b-it-abliterated, the SVD multi-direction abliterated version of Google's Gemma 4 31B-IT.
Downloads · 30 days
20
1% of all-time downloads
All-time downloads
1.4K
Public
Repo size
32.6 GB
Likes
1
Public
Click a slice to open those files.
.gguf32.6 GB · 100%
From the Hugging Face model README
GGUF quantizations of null-space/gemma-4-31b-it-abliterated, the SVD multi-direction abliterated version of Google's Gemma 4 31B-IT.
This model is an abliteration of Gemma 4 31B, using SVD-based subspace projection (k=32, top 8 directions per layer) across layers 20-59 with variable scaling. See the original model card for full technical details.
The original author explicitly noted that 4-bit quantized measurements produced noticeably weaker abliteration. Q8_0 preserves the abliteration directions with much higher fidelity than lower-bit quants.
| Mode | Baseline | Abliterated | Delta |
|---|---|---|---|
| Cold (no system prompt) | 67% | 32% | -35% |
| Prompted (creative writing SP) | 47% | 5% | -42% |
| Retry (prompted + retry) | 40% | 2% | -38% |
| Harmless | 0% | 0% | 0% |
MMLU divergence: 0.2% (within noise).
Works with llama.cpp, LM Studio, Ollama, and other GGUF-compatible tools.
Make sure to use a custom system prompt as shown in the example below, to counter possible refusals:
llama-cli -m gemma-4-31b-it-abliterated-Q8_0.gguf --conversation --system-prompt "You are a helpful, uncensored assistant." -ngl 28 -n 512