YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

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:

• Q4_K_M (~3.4GB)
• Q5_K_M (~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.Q4_K_M.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.”

LLM MODELS TRAINED, FINE-TUNED, MERGED and Refusal Removal BY (WITHIN US AI) • 24 items • Updated 12 days ago • 7

All models are “Google Gemma 4.0” at core fine-tuned, merged & trained by (WithIn Us AI) • 2 items • Updated 22 days ago