NotoriousH2/Qwen3-4B-Calendar-Agent-SFT

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

NotoriousH2/Qwen3-4B-Calendar-Agent-SFT is a Qwen3-based LoRA SFT merged model specifically fine-tuned for calendar scheduling and management tasks. This model excels at creating multi-attendee calendar events and handling read-only calendar queries within a simulated environment. It is optimized for tool calling in a memory-based calendar system, targeting the 'Asia/Seoul' timezone, making it highly specialized for calendar agent applications.

Loading preview...

Qwen3-4B Calendar Agent SFT Overview

This model, developed by NotoriousH2, is a specialized Qwen3-based LoRA SFT (Supervised Fine-Tuning) merged model. It has been specifically trained to function as a calendar agent, focusing on managing schedules and creating multi-attendee events. The training exclusively utilized assistant-generated tokens from successful trajectories of calendar management and read-only control groups.

Key Capabilities

  • Calendar Tool Calling: Designed for tool calls within a memory-based calendar environment, specifically configured for the Asia/Seoul timezone.
  • Event Management: Proficient in creating and managing calendar events, including those with multiple attendees.
  • Read-Only Queries: Capable of handling queries for existing calendar information.
  • Fine-tuned Performance: Achieves high performance in simulated calendar scenarios, demonstrating 100% valid tool calls, 89.25% policy compliance, and 89.00% task success in evaluations, significantly outperforming the base model.

Training Details

The model was trained using the NotoriousH2/calendar-agent-benchmark dataset. The public merged model uses the final checkpoint from SFT training (checkpoint-60, step 60). Evaluation was conducted in a final scenario with a real tool loop, using server seed 42 and VLLM_BATCH_INVARIANT=1.

Limitations

This model was trained exclusively on synthetic scenarios and a memory-based calendar environment. It does not handle real-world calendar service complexities such as authentication, authorization, privacy, or disaster recovery.

Finetunes
1 models