Model Overview

This model, choco800/qwen3-4b-agent-v11, is a 4 billion parameter language model fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model. It is provided as a fully merged model, eliminating the need to load a separate base model.

Key Capabilities

• Multi-turn Agent Task Performance: Specifically trained to enhance performance in multi-turn agent scenarios.
• Environment Interaction: Optimized for tasks requiring environment observation and action selection.
• Tool Use: Designed to effectively utilize tools within agent trajectories.
• Error Recovery: Capable of learning to recover from errors during complex tasks.
• Targeted Domains: Demonstrates improved performance on benchmarks like ALFWorld (household tasks) and DBBench (database operations).

Training Details

The model was fine-tuned using LoRA with Unsloth, with a maximum sequence length of 8192 tokens. Loss was applied exclusively to the assistant's responses, focusing the training on generating effective agent actions and observations. The training data includes u-10bei/dbbench_sft_dataset_react, which is distributed under the MIT License.