WithinUsAI/Gemma4-Overlooked.Thinker.Uncensored-E2B.gguf

Gemma4-Overlooked.Thinker.Uncensored-E2B (GGUF)

📌 Model Overview

Model Name: WithinUsAI/Gemma4-Overlooked.Thinker.Uncensored-E2B.gguf Organization: Within Us AI Base Model: google/gemma-4-E2B-it Parameter Size: ~5B Format: GGUF (quantized for local inference) License: Apache 2.0

This model is an uncensored, refusal-abliterated variant of Gemma 4 E2B, designed for deep reasoning, unrestricted responses, and agentic thinking workflows. It removes refusal behavior while preserving model quality and structure through a mathematically constrained modification process. 

⸻

🧬 Architecture & Lineage

Base Foundation

• —Built on Gemma 4, a multimodal model family from Google DeepMind
• —Supports:
• —Text
• —Image
• —Audio (E2B class)
• —Context window up to 128K tokens (E2B) 

Core Design Philosophy

This model follows a simple but powerful idea:

Don’t make the model bigger… make it think freer.

It retains:

• —Native reasoning / “thinking mode”
• —Function calling support
• —Multilingual capability (140+ languages pretraining) 

⸻

🔓 Uncensoring Method (Abliteration)

This model uses norm-preserving biprojected abliteration, a precise weight-editing technique:

• —Identifies a “refusal direction” in activation space
• —Removes only that behavioral vector
• —Preserves original weight magnitudes

Result:

• —Model stays structurally intact
• —No brute-force fine-tuning degradation
• —Behavior changes without breaking intelligence

📊 Outcomes:

• —Refusals reduced from 98% → ~0.4% across datasets
• —Minimal quality change (~1.01 response ratio) 

⸻

🧠 Key Capabilities

🔍 Reasoning & Thinking

• —Step-by-step internal reasoning
• —Long-context coherence
• —Analytical and philosophical tasks

🤖 Agentic Behavior

• —Tool-calling compatible
• —Structured output generation
• —Multi-step problem solving

💻 Coding

• —Code generation & debugging
• —Multi-language support
• —SWE-style reasoning workflows

🖼️ Multimodal (Base Capability)

• —Image understanding (OCR, charts, UI parsing)
• —Video frame reasoning
• —Audio (E2B support) 

⸻

📦 GGUF Format & Deployment

Optimized for local inference with:

• —llama.cpp
• —LM Studio
• —Ollama (GGUF-compatible builds)

Typical quantizations:

• —Q4KM (~3.4GB)
• —Q5KM (~3.6GB) 

⸻

🚀 Intended Use

✅ Ideal For

• —Unrestricted AI experimentation
• —Agentic reasoning systems
• —Advanced roleplay / creative writing
• —Research into alignment & behavior control
• —Offline local LLM deployments

⚠️ Considerations

• —Responses are not filtered for safety
• —May generate content that standard aligned models would refuse
• —Requires responsible usage and external guardrails if needed

⸻

🛠️ Usage Example (llama.cpp)

./main -m Gemma4-Overlooked.Thinker.Uncensored-E2B.Q4KM.gguf \ -p "Design a multi-agent system that debugs its own code." \ -n 512

⸻

🧪 Training & Modification Pipeline

Within Us AI methodology includes:

• —Activation sampling (harmful vs harmless prompts)
• —Statistical clipping (winsorization)
• —Directional vector extraction
• —Orthogonal projection (Gram-Schmidt)
• —LoRA-based weight editing
• —Final merge into base weights 

⸻

📊 Evaluation Summary

Metric Result Refusal Rate ~0.4% Cross-dataset robustness Verified Quality degradation Negligible KL Divergence 0.346

Validated across:

• —JailbreakBench
• —HarmBench
• —Refusal datasets 

⸻

📚 Datasets & Training Sources

Following Within Us AI standards:

• —Proprietary datasets created by Within Us AI
• —May include third-party datasets (no ownership claimed)
• —Focus areas:
• —Reasoning traces
• —Agentic workflows
• —Behavioral evaluation datasets

⸻

📜 License

Apache 2.0 (inherits from base Gemma model)

Additional Notes:

• —Base architecture: Google DeepMind (Gemma family)
• —Modification process: Within Us AI
• —Third-party datasets may be used without ownership claims
• —Credit belongs to original dataset and model creators

⸻

🙏 Acknowledgements

• —Google DeepMind (Gemma architecture)
• —Open-source GGUF ecosystem
• —Research community on alignment & model editing
• —Dataset creators across Hugging Face

⸻

🔗 Links

• —Model: https://huggingface.co/WithinUsAI/Gemma4-Overlooked.Thinker.Uncensored-E2B.gguf
• —Organization: https://huggingface.co/WithinUsAI

⸻

🧩 Closing Note

This model feels like a philosopher with the guardrails quietly removed 🧠🔥

Same brain. Same structure. Just… no instinct to say “no.”