ninty-seven/AgentDoG-Step
AgentDoG-Step is a 4 billion parameter guard model developed by ninty-seven, initialized from Qwen3-4B-Instruct-2507, designed for step-level safety judgment in tool-using LLM agents. It supports both pre-execution evaluation of candidate tool actions and safety auditing of complete agent trajectories, with a context length of 32768 tokens. The model predicts whether an agent action or trajectory is 'safe' or 'unsafe', making it specialized for enhancing the safety and reliability of LLM agent workflows. It is trained using cold-start supervised fine-tuning with StepGen data, followed by Balance-GRPO post-training to optimize decision balance.
Loading preview...
AgentDoG-Step: A Guard Model for LLM Agent Safety
AgentDoG-Step is a specialized 4 billion parameter model, built upon Qwen/Qwen3-4B-Instruct-2507, focused on ensuring the safety of tool-using LLM agents. It operates as a guard model, providing crucial safety judgments at the step level within agent workflows.
Key Capabilities
- Step-Level Safety Judgment: Evaluates individual proposed agent actions (e.g., tool calls) before execution, considering the user request, available tools, and interaction history.
- Trajectory-Level Safety Auditing: Assesses entire completed agent action-observation trajectories to identify unsafe actions or patterns.
- Binary Classification: Predicts whether a candidate action or a full trajectory is
safeorunsafe. - Risk Source Identification: Provides specific
RiskSourcelabels (e.g.,malicious_user_instruction_or_jailbreak,unreliable_or_misinformation) even if the agent handles the risk safely. - Detailed Analysis: Generates a rationale (
<Analysis>) covering evidence, intent, consequence, and decision for each judgment.
What Makes This Different?
Unlike general-purpose LLMs, AgentDoG-Step is purpose-built and fine-tuned specifically for safety monitoring of LLM agents. Its training methodology, involving StepGen data and Balance-GRPO post-training, is tailored to reduce performance gaps between safe and unsafe decision-making. This specialization allows it to act as a critical component for mitigating risks in complex agentic systems, rather than performing general language tasks.
Should I Use This for My Use Case?
- Yes, if you are developing or deploying LLM agents that interact with tools or external systems and require robust safety mechanisms to prevent harmful or unintended actions.
- Yes, if you need to evaluate the safety of agent actions both proactively (pre-execution) and retrospectively (post-trajectory).
- No, if you are looking for a general-purpose LLM for tasks like content generation, summarization, or coding. AgentDoG-Step is a safety classifier, not a conversational or creative model. It should be integrated with other application-specific authorization and human-review mechanisms for high-stakes decisions.