Gemma-4-E4B-it-ultra-uncensored-heretic-pruned-TextOnly-EnglishOnly-GGUF

license: apache-2.0 language:

en base_model:
llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic tags:
EnglishOnly
TextOnly
NSFW
Uncensored
Decensored
Heretic
ARA
VisionRemoved
AudioRemoved
Gemma4
Gemma

After thorough usage and testing, I chose two UNCENSORED Gemma 4 E4B Models to be pruned (converted to English and Text-to-Text only) so I can use them on my potato laptop/phone.

The Source Models

1. Gemma-4-E4B-it-uncensored

Profile: The most uncensored G4-E4B model.
Stats: 0/656 manually checked refusals (according to the author). KL divergence value of 0.068 (very negligible).

2. Gemma-4-E4B-it-ultra-uncensored-heretic (Base for this pruned version)

Profile: The most intelligent uncensored G4-E4B model with the lowest KL divergence on Hugging Face.
Stats: KL divergence of 0.0076. Refusals 3/100 (which you probably won't notice anyways).

This pruned model is the more intelligent one, while still being really heavily uncensored.

Check out the other one here: Gemma-4-E4B-it-uncensored-pruned-Text-and-English-ONLY-GGUF.

Modifications

What's Changed:

Vision & Audio Modules: REMOVED (Just Text-to-Text now).
Language Tokens: ~61,000 (23.1%) REMOVED This includes non-Latin, Cyrillic, Arabic, Asian languages, etc.
- Note: Languages using some Latin letters like German and Vietnamese are still there. Latin was not removed to keep the logic alive.
Weights: Cut down from >2300 to 720 (loads fast as fahhhh).

What DID NOT Change:

Layers
Logic (as far as I tested)
Context
No training/retraining done

TL;DR: Overall, both pruned models are faster and significantly less resource-intensive.

Base Model Usage Settings

You might want to change these according to your preferences, but here is the baseline:

Temperature: 1.0
Top P: 0.95
Top K: 64
Jinja --jinja (don't ignore this one, especially on llama.cpp!)

Colab Notebook I made and used to prune these models:

Gemma-4-E4B Pruner

Thanks:

InfinimindCreations
llmfan46
Claude.ai (for coding the notebook)
Kaggle.com (for the free resources)

These were my first attempts at anything to do with modifying a model so if you wanna suggest something, do it.

The Source Models

Profile: The most uncensored G4-E4B model.

Stats: 0/656 manually checked refusals (according to the author). KL divergence value of 0.068 (very negligible).

Profile: The most intelligent uncensored G4-E4B model with the lowest KL divergence on Hugging Face.

Stats: KL divergence of 0.0076. Refusals 3/100 (which you probably won't notice anyways).

This pruned model is the more intelligent one, while still being really heavily uncensored.

Check out the other one here: Gemma-4-E4B-it-uncensored-pruned-Text-and-English-ONLY-GGUF.

Modifications

What's Changed:

Vision & Audio Modules: REMOVED (Just Text-to-Text now).

Language Tokens: ~61,000 (23.1%) REMOVED This includes non-Latin, Cyrillic, Arabic, Asian languages, etc.

Note: Languages using some Latin letters like German and Vietnamese are still there. Latin was not removed to keep the logic alive.

Weights: Cut down from >2300 to 720 (loads fast as fahhhh).

What DID NOT Change:

Layers

Logic (as far as I tested)

Context

No training/retraining done

TL;DR: Overall, both pruned models are faster and significantly less resource-intensive.