parth-1/metaguard-policy-agent-v1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Apr 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The MetaGuard Ad Policy Agent v1, developed by parth-1, is an experimental ad policy compliance agent fine-tuned using Group Relative Policy Optimization (GRPO) via Unsloth and Hugging Face's TRL library. This model is specifically designed to operate within the MetaGuard Ad Policy Sandbox, a multi-agent reinforcement learning environment. It focuses on gathering signals and making terminal enforcement decisions related to ad policy compliance. The agent's primary strength lies in its information-gathering phase, though it currently struggles with strict terminal sequence chaining.

Loading preview...

MetaGuard Ad Policy Agent (v1) Overview

This model, developed by parth-1, is an experimental ad policy compliance agent. It has been fine-tuned using Group Relative Policy Optimization (GRPO) with Unsloth and Hugging Face's TRL library, specifically for the MetaGuard Ad Policy Sandbox environment. This multi-agent reinforcement learning setup requires the model to gather signals and make enforcement decisions.

Key Characteristics & Performance

  • Training: Fine-tuned using GRPO, which significantly improved the agent's intermediate signal gathering phase, increasing the mean reward per step from -0.185 to -0.050 compared to the baseline.
  • Environment Interaction: Designed to navigate a multi-agent reinforcement learning environment, focusing on ad policy compliance.
  • Strengths: Shows strong capability in the information-gathering phase, such as calling query_regulations and analyze_image.

Critical Usage Information

  • JSON Schema Errors (422): When interacting with the MetaGuard API server, the model's output may cause 422 Unprocessable Entity errors due to hallucinated, invalid tracking fields in its JSON. A provided Python parser patch is required to sanitize the output, ensuring strict JSON validation and preventing API crashes.

Limitations

  • The agent currently struggles with strict terminal sequence chaining, frequently failing to call the submit_audit action before attempting final approve or reject decisions, leading to incomplete task execution.

Finetunes
1 models