NotoriousH2/Qwen3-4B-Calendar-Agent-RLVR
NotoriousH2/Qwen3-4B-Calendar-Agent-RLVR is a 4-billion parameter merged model developed by NotoriousH2, specifically designed for Korean calendar management tool calling. It was trained using GRPO by evaluating the execution trajectories and final states of calendar tools as rewards. This model excels at managing calendar events, including 조회 (inquiry), 생성 (creation), 수정 (modification), and 삭제 (deletion), and handles complex scenarios like conflict resolution and finding common available times for multiple attendees. Its primary use case is to act as an intelligent agent for automating and streamlining calendar-related tasks in Korean.
Loading preview...
Overview
NotoriousH2/Qwen3-4B-Calendar-Agent-RLVR is a 4-billion parameter merged model, fine-tuned for Korean calendar management tool calling. It was trained using Grouped Reward Policy Optimization (GRPO), where the execution trajectories and final states of calendar tools were used as rewards. This approach allows the model to learn complex sequences of actions for effective calendar interaction.
Key Capabilities
- Comprehensive Calendar Management: Supports inquiry, creation, modification, and deletion of calendar events.
- Intelligent Information Handling: Capable of requesting necessary information from the user when details are missing and confirming user approval before deletion.
- Conflict Resolution: Identifies schedule conflicts and can generate available time slots.
- Multi-Attendee Scheduling: Can explore common available times for multiple participants.
- Robust Tool Calling: Achieved 99.00% Task Success and 99.25% Policy Compliance in evaluations, significantly outperforming its SFT counterpart.
Training and Evaluation
The model's training incorporated a unique reward system, evaluating safety rules, procedural progress, and final success. It was evaluated on 400 synthetic scenarios, demonstrating its ability to convert 41 failed cases from the SFT model into successes. While highly effective, it's noted that the model was trained and evaluated in a synthetic Asia/Seoul timezone environment and does not cover real-world aspects like authentication, permissions, privacy, or network failures. The "conflict after creation" template showed slightly lower performance (76/80) compared to others.