Overview

This repository provides a LoRA adapter for the Qwen/Qwen3-4B-Instruct-2507 base model, developed by Amouri28. The adapter, weighing 4 billion parameters, is fine-tuned using LoRA + Unsloth to significantly improve performance on multi-turn agent tasks.

Key Capabilities

• Enhanced Multi-Turn Agent Performance: Specifically trained to excel in complex, multi-turn interactions.
• Task Specialization: Optimized for household tasks (ALFWorld) and database operations (DBBench).
• Comprehensive Learning: The training objective applies loss to all assistant turns, fostering better environment observation, action selection, tool use, and error recovery.
• Efficient Fine-tuning: Utilizes LoRA (full precision base) with a max sequence length of 2048 and a learning rate of 2e-06.

Good For

• Developers working on agent-based systems requiring robust multi-turn interaction capabilities.
• Applications involving automated household tasks or complex database operations.
• Extending the Qwen3-4B-Instruct-2507 model's functionality for specialized agentic workflows. Users must comply with the MIT License of the training data and the base model's original terms.