Qwen3-4B ALFWorld+DBBench Mixed LoRA Adapter

This repository provides a LoRA adapter (r=64, alpha=128) fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model. It is specifically designed to improve the base model's performance on complex, multi-turn agent tasks.

Key Capabilities

• Enhanced Multi-Turn Agent Performance: Optimized for tasks requiring sequential decision-making and interaction.
• Specialized for ALFWorld: Improves the model's ability to handle household tasks, including environment observation and action selection.
• Specialized for DBBench: Enhances performance in database operation tasks, including tool use and error recovery.
• Trajectory-Based Learning: Loss is applied to all assistant turns in a multi-turn trajectory, allowing the model to learn from complete interaction sequences.

Good For

• Developing agents that need to perform complex, multi-step tasks in simulated environments.
• Applications requiring improved tool use and error recovery in agentic workflows.
• Researchers and developers working on AI agents for household automation or database interaction.
• Extending the capabilities of the Qwen3-4B-Instruct-2507 base model for specific agentic use cases.