Overview

This repository provides a LoRA adapter (r=64, alpha=128) fine-tuned by thetmon from the Qwen/Qwen3-4B-Instruct-2507 base model using LoRA + Unsloth. It contains only the adapter weights, requiring the base model to be loaded separately.

Key Capabilities

• Multi-turn Agent Task Performance: Specifically trained to improve performance in complex, multi-step agent tasks.
• Environment Interaction: Learns to process environment observations and select appropriate actions.
• Tool Use: Enhanced capabilities for integrating and utilizing tools within task trajectories.
• Error Recovery: Designed to improve the model's ability to recover from errors during multi-turn interactions.

Training Details

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA (full precision base)
• Max Sequence Length: 4096 tokens
• Epochs: 3
• Learning Rate: 2e-04
• Training Data: Utilizes u-10bei/sft_alfworld_trajectory_dataset_v5 and u-10bei/dbbench_sft_dataset_react_v4, both under the MIT License.

Good For

• Agentic Workflows: Ideal for applications requiring an AI agent to perform sequential tasks.
• ALFWorld Tasks: Excels in household task automation scenarios.
• DBBench Operations: Suited for database interaction and operation tasks.