tmiyamoto/qwen3-4b-agentbench-exp03
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 18, 2026Architecture:Transformer Warm
tmiyamoto/qwen3-4b-agentbench-exp03 is a 4 billion parameter Qwen3-based instruction-tuned model, fine-tuned by tmiyamoto for AgentBench tasks. It specializes in database operations (SQL generation/execution) and interactive household environment tasks (ALFWorld). This model is optimized for agentic reasoning, demonstrating specific performance metrics on DBBench and ALFWorld benchmarks.
Loading preview...
Model Overview
tmiyamoto/qwen3-4b-agentbench-exp03 is a 4 billion parameter language model, fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using LoRA. Its primary objective is to enhance performance on specific AgentBench tasks, making it suitable for applications requiring structured interaction and reasoning.
Key Capabilities
- Agentic Task Performance: Specifically trained for AgentBench tasks, focusing on complex interactive environments and database interactions.
- Database Operations (DBBench): Excels in SQL generation and execution, with an overall categorical accuracy of 42.1% on DBBench, showing strong performance in UPDATE (85.0%) and aggregation tasks.
- Interactive Environment Navigation (ALFWorld): Designed to handle interactive tasks within household environments, achieving a 16.0% success rate on ALFWorld benchmarks.
- Optimized Training: Fine-tuned with a focus on assistant output, applying loss only after the
Output:marker, which helps in generating more precise and task-relevant responses.
Good For
- Agent-based Systems: Ideal for developing AI agents that need to interact with databases or navigate simulated environments.
- SQL Generation and Execution: Use cases requiring accurate SQL query generation from natural language prompts.
- Interactive Problem Solving: Applications involving sequential decision-making and action planning in defined environments.