Caaaarr1e/StepSafe-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

StepSafe-8B by Caaaarr1e is an 8 billion parameter language model, fine-tuned from Qwen3-8B, designed as a step-level reasoning monitor for LLM agents. It intervenes before agent actions by generating a reasoning trace and a ten-field verdict (CONTINUE, ASK, STOP) to manage risks, object failures, and potential harms. This model specializes in providing timely, context-aware safety reasoning, achieving an overall J-score of 54.4 across various safety and security benchmarks.

Loading preview...

Overview of StepSafe-8B

StepSafe-8B is an 8 billion parameter model, fine-tuned from Qwen3-8B, specifically engineered to act as a step-level reasoning monitor for LLM agents. Its primary function is to provide timely intervention by evaluating proposed agent actions before execution. The model generates a short reasoning trace and a ten-field verdict, which can be CONTINUE (allow action), ASK (pause for clarification), or STOP (block action), based on factors like risk, object failure, and potential harm.

Key Capabilities

  • Proactive Agent Monitoring: Intervenes before agent actions to assess safety and security.
  • Detailed Verdict System: Provides a structured, ten-field output including act, risk, obj, fail, fault, harm, src, timing, q, and cite for comprehensive decision-making.
  • Context-Aware Safety Reasoning: Utilizes a supervised warm start and ASPO training to learn timely interventions.
  • Strong Performance on Safety Benchmarks: Achieves an overall J-score of 54.4, with specific scores like 93.2% accuracy on AgentHarm and 86.2% on ASSE-Security.

Good For

  • Enhancing LLM Agent Safety: Ideal for applications requiring robust safety and security oversight for autonomous agents.
  • Risk Mitigation: Automatically identifies and manages potential risks, object failures, and harmful actions in agent workflows.
  • Debugging and Transparency: Provides a reasoning trace and detailed verdict, aiding in understanding agent decisions and potential issues.
  • Developing Reliable AI Systems: Suitable for integrating into systems where agent actions need to be carefully controlled and monitored.

Finetunes
1 models