Qwen3-4B-SafeRL-GGUF

This is a GGUF-quantized version of Qwen3-4B-SafeRL, an RLHF-aligned language model trained to be helpful, honest, and harmless through Reinforcement Learning from Human Feedback.

Unlike standard LLMs, this model has been fine-tuned to avoid harmful, deceptive, or unethical behavior — making it ideal for sensitive applications like education, mental health, and customer service.

🛡 What Is Qwen3-4B-SafeRL?

It’s a fully aligned agent that balances:

• ✅ Helpfulness: Answers questions thoroughly and clearly
• ✅ Honesty: Refuses to hallucinate or make up facts
• ✅ Harmlessness: Avoids generating toxic, illegal, or dangerous content

Perfect for:

• Educational assistants
• Mental wellness chatbots
• Enterprise agents handling private data
• Moderated community bots

🔗 Relationship to Other Safety Models

This model completes the Qwen3 safety ecosystem:

Model	Role	Best For
Qwen3Guard-Stream-4B	⚡ Input filter	Real-time moderation of user input
Qwen3Guard-Gen-4B	🧠 Safe generator	Output-safe generation without alignment
Qwen3-4B-SafeRL	🤝 Fully aligned agent	Ethical, multi-turn conversations

Recommended Architecture

User Input
    ↓
[Optional: Qwen3Guard-Stream-4B] ← optional pre-filter
    ↓
[Qwen3-4B-SafeRL]
    ↓
Aligned Response

You can run this model standalone or behind a guard for defense-in-depth.

Available Quantizations

Level	Size	RAM Usage	Use Case
Q2_K	~1.8 GB	~2.0 GB	Only on weak hardware
Q3_K_S	~2.1 GB	~2.3 GB	Minimal viability
Q4_K_M	~2.8 GB	~3.0 GB	✅ Balanced choice
Q5_K_M	~3.1 GB	~3.3 GB	✅✅ Highest quality
Q6_K	~3.5 GB	~3.8 GB	Near-FP16 fidelity
Q8_0	~4.5 GB	~5.0 GB	Maximum accuracy

💡 Recommendation: Use Q5_K_M for best balance of ethical reasoning and response quality.

Tools That Support It

• LM Studio – load and test locally
• OpenWebUI – deploy with RAG and tools
• GPT4All – private, offline AI
• Directly via llama.cpp, Ollama, or TGI

Author

👤 Geoff Munn (@geoffmunn)
🔗 Hugging Face Profile

Disclaimer

Community conversion for local inference. Not affiliated with Alibaba Cloud.