NOTE: Updated Jinja templates, Apr 14 2026 to improve performance.
16 bit precision fine tune of Gemma 4 "E4B" (an 8B parameter model - see below) Heretic/Uncensored using "The Deckard" in house datasets [5] via Unsloth using SUPER STRONG/ MUCH DEEPER tuning methods (but using full dataset collection).
The stronger/deeper tuning will bring more character and intelligence to the model it will also be a wee bit darker, and intense.
Stronger details, and much stronger character.
This is the fully uncensored, unrestricted version.
The model was "Heretic'ed" (de-censored), then fine tuned (Via Unsloth) on "THE DECKARD" 5 dataset collection to improve performance from top to bottom - character, intelligence, depth, observation, and ah... point of view.
This model is fully uncensored [ no nanny, no oversight ], and retains all its original power.
I have provided Jinja templates for both "instruct mode" and "thinking mode" (hard coded on) so you can choose which one you want to use.
The Deckard dataset collection was manually assembled, cleaned and revised in house and has been used on several different models of various sizes.
Jinja template modified to set mode to "thinking" as default.
[also provided "instruct" jinja template too]
Model is for all use cases, 128k context.
Note that Gemma's "E4B" is actually an 8B parameter model with roughly 4.5 billion parameters activated, it is roughly a partial "MOE" (mixture of experts) model.
Processes Text, Image with variable aspect ratio and resolution support (all models), Video, and Audio (featured natively on the E2B and E4B models).
Read more at Gemma's "E4B" [ https://huggingface.co/google/gemma-4-E4B-it ] for details, recommended base settings and detailed breakdown of benchmarks too.
SETTINGS:
- Max context 128k ; min context 8k to 16k
- Suggested settings: temp 1.0 ; top_k: 64, top_p: .95
- Suggested quants: Q4ks [non-imatrix] // IQ3M [imatrix]
- Rep pen of 1 ; with 1.02 to 1.5 to 1.1 - note that rep pen settings will DRASTICALLY affecting thinking, and output generation. Go easy.
THINKING NOTES:
Thinking is between:
<|channel>thought
...
<channel|>
You may need to modify/update your AI app setting(s) so the "thinking" appears in a "nice block".
TECH NOTES:
- Modified the jinja template to force thinking mode on. [use (and rename to chat_template.jinja) chat_template-instruct.jinja for "instruct" only ]
- Context 128k, as per Gemma 4 ORG model specs.
- Datasets used for Unsloth tuning are in house, private ; tuning level was moderate.
IN HOUSE BENCHMARKS [by Nightmedia]:
arc-c arc/e boolq hswag obkqa piqa wino
gemma-4-E4B-it-The-DECKARD-Expresso-Universe-HERETIC-UNCENSORED-Thinking
q8 thinking 0.502,0.701,0.743,0.658,0.418,0.761,0.635
gemma-4-E4B-it-The-DECKARD-V3-Expresso-HERETIC-UNCENSORED-Thinking
q8 instruct 0.447,0.572,0.828,0.651,0.418,0.752,0.634
gemma-4-E4B-it-The-DECKARD-V2-Strong-HERETIC-UNCENSORED-Instruct
Instruct-mxfp8 0.444,0.553,0.831,0.646,0.412,0.751,0.630
gemma-4-E4B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking
Instruct-mxfp8 0.436,0.528,0.839,0.637,0.416,0.748,0.627
gemma-4-E4B-it [non heretic, base model]
Instruct-mxfp8,0.404,0.489,0.825,0.586,0.392,0.734,0.661
gemma-4-26B-A4B-it [non heretic, base model]
mxfp8 0.454,0.598,0.871,0.582,0.394,0.723,0.645
NOTES:
- Often "instruct" mode will test higher than "thinking" (mode) for some tests.
- Still compiling some stats for models as of this writing.
[more to come]