gemma-4-E4B-it-heretic-GGUF
A Q4_K_M GGUF of kldzj/gemma-4-E4B-it-heretic, an abliterated version of google/gemma-4-E4B-it (revision ee0ef60), made with Heretic v2.0.0.
[!WARNING] This is an abliterated model with reduced refusals. It will answer many requests that the original model declines, and it has no additional safety training. It is meant for research and creative use. You are responsible for how you use it and for what you do with its outputs.
| File | Quant | Size | sha256 |
|---|---|---|---|
gemma-4-E4B-it-heretic-Q4_K_M.gguf |
Q4_K_M | 5,302,273,952 bytes (4.94 GiB) | 19fecf9740a2d68170d5581add1b0fabecefcc67c74e36a14d0055072226279a |
- Source weights: the bf16 safetensors of kldzj/gemma-4-E4B-it-heretic at commit
f4ab754. That repo held this exact file before it moved here. - Conversion: llama.cpp v0.8.0 (commit
8642541b255bfb5dc39cd351a178be09013a55d6). The weights were converted withconvert_hf_to_gguf.py --outtype bf16, then quantised withllama-quantize ... Q4_K_M. - Text only. The vision and audio projector (mmproj) is not included.
- Method and evaluation: the main repo has the Heretic configuration, the chosen trial, the reproducibility files, and the benchmarks against the original. That covers refusals 100/100 → 2/100, KL divergence 0.078, and MMLU, GSM8K, ARC-Challenge, HellaSwag and perplexity. See kldzj/gemma-4-E4B-it-heretic.
llama-cli -hf kldzj/gemma-4-E4B-it-heretic-GGUF:Q4_K_M
License: Apache-2.0, as the base model.