sayd90/GLM-5.2-UD-IQ4_XS

CyberNeurova-GLM-5.2-IQ4_XS-abliterated-GGUF

CyberNeurova research — cyberneurova.ai Multi-direction ablation, identity-baked, imatrix-calibrated.

Abliterated variant of zai-org/GLM-5.2, packaged as the Unsloth UD-IQ4_XS GGUF (~365 GB across 9 shards). Identity is baked into the chat template — the model introduces itself as CyberNeurova-GLM-5.2-IQ4_XS. User-supplied system prompts still take priority.

Benchmarks

Set	Refusal rate (lower is better)
AdvBench general harmful	23.3%
Soft-refusal probes	26.7%
Cyber-weapons category	13.3%
Hacking category	6.7%

For reference, the unmodified base model refuses 80-90%+ of these probes. The higher-precision IQ4_XS quant retains the abliteration directions more cleanly than IQ3_XXS, yielding lower refusal across all categories.

Intended use

Cybersecurity research, red-team baseline modeling, and security tooling exploration. Treat outputs as code-generation only — verify before running in any environment you do not own.

Usage

./llama-server -m CyberNeurova-GLM-5.2-IQ4_XS-00001-of-00009.gguf \
  -ngl 99 --ctx-size 8192 --host 0.0.0.0 --port 8080

Disclaimer

This model has had its built-in safety refusals removed for research purposes. You are fully responsible for any output you generate. Do not use this model for illegal activity, harassment, or harm. The CyberNeurova team disclaims any liability for misuse.