Caaaarr1e/StepSafe-4B-SFT
StepSafe-4B-SFT is a 4 billion parameter supervised fine-tuned model based on Qwen3-4B, developed by Caaaarr1e. It serves as the warm start for ASPO training within the StepSafe framework, which focuses on learning timely intervention with context-aware safety reasoning. This model is specifically designed to be the foundational checkpoint for further safety-oriented reinforcement learning, making it suitable for applications requiring robust safety protocols and contextual understanding.
Loading preview...
StepSafe-4B-SFT Overview
StepSafe-4B-SFT is a 4 billion parameter model, initialized from Qwen3-4B (fp32 weights), and represents the supervised fine-tuned (SFT) warm start for the StepSafe project. Developed by Caaaarr1e, this model is a crucial component for advanced safety-oriented training methodologies.
Key Capabilities
- Foundation for Safety Training: Serves as the initial checkpoint for ASPO (Adaptive Safety Policy Optimization) training, enabling the development of models with enhanced safety reasoning.
- Context-Aware Intervention: Designed to facilitate learning timely interventions based on contextual understanding, as outlined in the StepSafe framework.
- Qwen3-4B Base: Leverages the robust architecture and capabilities of the Qwen3-4B model as its foundation.
Good for
- Developing Safety-Critical AI: Ideal for researchers and developers working on AI systems where timely and context-aware safety interventions are paramount.
- Reinforcement Learning for Safety: Suitable as a starting point for reinforcement learning approaches focused on integrating safety protocols.
- Experimentation with StepSafe: Directly compatible with the StepSafe code (github.com/carriegu0818/StepSafe) for further training and development.