Overview

This repository provides a LoRA adapter fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model, utilizing LoRA + Unsloth for efficient training. It contains only the adapter weights, requiring the base model to be loaded separately.

Key Capabilities

• Enhanced Multi-Turn Agent Performance: Specifically trained to improve the model's ability to handle complex, multi-turn agent tasks.
• Task Domains: Optimized for performance in ALFWorld (household tasks) and DBBench (database operations).
• Trajectory Learning: The training objective applies loss to all assistant turns in a multi-turn trajectory, enabling the model to learn:
  • Environment observation
  • Action selection
  • Tool use
  • Error recovery

Training Details

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA (full precision base)
• Max Sequence Length: 2048 tokens
• Epochs: 15
• Learning Rate: 2e-06
• LoRA Configuration: r=64, alpha=128
• Training Data: Utilizes datasets such as u-10bei/dbbench_sft_dataset_react_v3 and u-10bei/sft_alfworld_trajectory_dataset_v5, both distributed under the MIT License.

Good For

• Developers building autonomous agents that require robust multi-turn interaction and task completion.
• Applications involving complex environments where agents need to observe, act, use tools, and recover from mistakes.
• Research into agentic LLM capabilities and trajectory-based learning.