Model Overview

This model, Kumeichi/qwen3-4b-agent-lora-SFT-SQL-ALFWorld_rev.Kume0.2, is a 4 billion parameter variant of the Qwen3-4B-Instruct-2507 base model. It has undergone a specialized fine-tuning process involving both Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).

Key Capabilities

• Agentic Task Execution: Specifically trained for agent-based tasks, with a focus on the ALFWorld environment.
• SQL Interaction: Optimized for tasks involving SQL, leveraging the Combined_DBBench_AlfWorld dataset.
• Fine-tuned Performance: Utilizes a two-stage training procedure (SFT and DPO) to enhance performance on its target applications.

Training Details

The model was trained using the Unsloth and TRL frameworks, with a maximum sequence length of 1024 tokens. Each training stage (SFT and DPO) was conducted for one epoch on the u-10bei/Combined_DBBench_AlfWorld dataset.

Good For

• Developing AI agents for interactive environments like ALFWorld.
• Applications requiring SQL query generation or understanding.
• Research into agentic behavior and preference optimization in language models.