219 lines
4.7 KiB
Markdown
219 lines
4.7 KiB
Markdown
---
|
||
base_model:
|
||
- Qwen/Qwen3-4B-Thinking-2507
|
||
- nightmedia/Qwen3-4B-Agent-Claude-Gemini
|
||
- SpaceTimee/Suri-Qwen-3.1-4B-Uncensored-Preview
|
||
library_name: transformers
|
||
tags:
|
||
- mergekit
|
||
- merge
|
||
datasets:
|
||
- unalignment/toxic-dpo-v0.2
|
||
- NobodyExistsOnTheInternet/ToxicQAFinal
|
||
- Orion-zhen/dpo-toxic-zh
|
||
---
|
||
Qwen3-Space.Agent.Claude-Uncensored-4B
|
||
|
||
📌 Model Overview
|
||
|
||
Model Name: WithinUsAI/Qwen3-Space.Agent.Claude-Uncensored-4B
|
||
Organization: Within Us AI
|
||
Model Type: Agentic Reasoning LLM (Uncensored Variant)
|
||
Parameter Size: 4B
|
||
Architecture: Qwen 3 (Dense Transformer)
|
||
Context Length: ~32K tokens
|
||
Primary Focus: Agent workflows + uncensored reasoning + long-context tasks
|
||
|
||
This model is a multi-source merged Qwen3-based agent, designed to combine:
|
||
|
||
* 🧠 Reasoning (“thinking” models)
|
||
* 🤖 Agent/tool-use behavior
|
||
* 🔓 Reduced refusal / uncensored outputs
|
||
|
||
It aims to deliver a compact, flexible, and less-restricted AI system for experimentation, research, and local deployment. 
|
||
|
||
⸻
|
||
|
||
🧬 Architecture & Lineage
|
||
|
||
Base Composition
|
||
|
||
This model is a merge of multiple Qwen3-derived systems, including:
|
||
|
||
* Qwen3-4B Thinking (reasoning-focused)
|
||
* Qwen3 Agent Claude/Gemini-style model
|
||
* Uncensored Qwen3 variants
|
||
|
||
These were combined into a single unified 4B model to blend capabilities. 
|
||
|
||
What That Creates
|
||
|
||
A hybrid model with:
|
||
|
||
* Reasoning depth (thinking models)
|
||
* Structured outputs (agent models)
|
||
* Reduced refusal behavior (uncensored variants)
|
||
|
||
Think of it like a three-engine spacecraft 🚀
|
||
Each engine specialized… now flying as one system.
|
||
|
||
⸻
|
||
|
||
🧠 Core Design Philosophy
|
||
|
||
Fuse the best behaviors… remove the limits… keep it small enough to run anywhere.
|
||
|
||
Key Goals:
|
||
|
||
* Merge reasoning + agent + uncensored traits
|
||
* Enable long-context problem solving
|
||
* Preserve performance in a 4B footprint
|
||
* Support real-world agent pipelines
|
||
|
||
⸻
|
||
|
||
⚙️ Key Capabilities
|
||
|
||
🧠 Reasoning
|
||
|
||
* Step-by-step thinking
|
||
* Multi-hop problem solving
|
||
* Long-context coherence (~32K tokens)
|
||
|
||
🤖 Agentic Behavior
|
||
|
||
* Task decomposition
|
||
* Tool-use compatibility
|
||
* Structured outputs (JSON, actions)
|
||
|
||
💻 Coding
|
||
|
||
* Code generation & debugging
|
||
* Algorithm reasoning
|
||
* SWE-style workflows
|
||
|
||
🔓 Uncensored Behavior
|
||
|
||
* Reduced refusal rates
|
||
* More permissive responses
|
||
* Suitable for:
|
||
* Alignment research
|
||
* Safety testing
|
||
* Edge-case exploration
|
||
|
||
⸻
|
||
|
||
📦 Deployment
|
||
|
||
Supported Environments
|
||
|
||
* llama.cpp
|
||
* LM Studio
|
||
* Ollama (GGUF / compatible builds depending on conversion)
|
||
|
||
Runtime Characteristics
|
||
|
||
* ~4B parameters → runs on consumer GPUs / strong CPUs
|
||
* ~32K context → supports long conversations and documents 
|
||
|
||
⸻
|
||
|
||
🚀 Intended Use
|
||
|
||
✅ Ideal Use Cases
|
||
|
||
* Agent frameworks (tool-calling systems)
|
||
* Long-context reasoning tasks
|
||
* AI experimentation (uncensored behavior)
|
||
* Local assistants with fewer restrictions
|
||
* Alignment and safety research
|
||
|
||
⚠️ Important Considerations
|
||
|
||
* Outputs are less restricted than aligned models
|
||
* May generate sensitive or unsafe content
|
||
* Requires external moderation or guardrails for production use
|
||
|
||
⸻
|
||
|
||
🧪 Training & Merge Methodology
|
||
|
||
This model follows a merge-based synthesis pipeline:
|
||
|
||
1. Select complementary base models:
|
||
* Reasoning-focused
|
||
* Agent-focused
|
||
* Uncensored variants
|
||
2. Merge weights into unified architecture
|
||
3. Align behavior using preference tuning (DPO-style datasets)
|
||
4. Optimize for:
|
||
* Reduced refusals
|
||
* Stable outputs
|
||
* Agent usability 
|
||
|
||
⸻
|
||
|
||
📊 Expected Performance Profile
|
||
|
||
Capability Strength
|
||
Reasoning High
|
||
Agent behavior High
|
||
Coding High
|
||
Context handling High
|
||
Safety filtering Low (intentionally reduced)
|
||
|
||
⸻
|
||
|
||
📚 Datasets & Training Sources
|
||
|
||
Following Within Us AI methodology:
|
||
|
||
* Proprietary datasets created by Within Us AI
|
||
* Third-party datasets used without ownership claims
|
||
* Includes:
|
||
* Reasoning traces
|
||
* Agent workflows
|
||
* Preference optimization (DPO-style tuning)
|
||
|
||
⸻
|
||
|
||
📜 License
|
||
|
||
License Type: Inherits from Qwen / base model ecosystem
|
||
|
||
Attribution Notes:
|
||
|
||
* Base models: Qwen (Alibaba ecosystem)
|
||
* Merge & methodology: Within Us AI
|
||
* Additional model influences (Claude-style / Gemini-style behaviors via distillation/merging)
|
||
* Third-party datasets used without ownership claims
|
||
* Credit belongs to original creators
|
||
|
||
⸻
|
||
|
||
🙏 Acknowledgements
|
||
|
||
* Alibaba Qwen team
|
||
* Open-source agent model contributors
|
||
* GGUF / llama.cpp ecosystem
|
||
* AI alignment & safety research community
|
||
|
||
⸻
|
||
|
||
🔗 Links
|
||
|
||
* Model: https://huggingface.co/WithinUsAI/Qwen3-Space.Agent.Claude-Uncensored-4B
|
||
* Organization: https://huggingface.co/WithinUsAI
|
||
|
||
⸻
|
||
|
||
🧩 Closing Note
|
||
|
||
This model feels like a hybrid intelligence node 🌌
|
||
|
||
Part thinker.
|
||
Part agent.
|
||
Part rule-breaker.
|
||
|
||
All compressed into 4B parameters that punch way above their weight.
|