Model Overview

The choco800/qwen3-4b-agent-v17 is a 4 billion parameter language model, fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model. This version is provided as a fully merged model, meaning it includes all weights and does not require loading a separate base model, simplifying deployment.

Key Capabilities

• Enhanced Multi-Turn Agent Performance: Specifically trained to improve performance in complex, multi-turn agent tasks.
• Agentic Trajectory Learning: Optimized for learning from assistant turns, covering environment observation, action selection, tool use, and error recovery.
• ALFWorld Optimization: Demonstrates improved capabilities in household task environments like ALFWorld.
• Efficient Training: Utilizes LoRA and Unsloth for efficient fine-tuning, with loss applied to all assistant turns in the trajectory.

Training Details

The model was trained for 1 epoch with a max sequence length of 8192, using a learning rate of 3e-06. It incorporates NEFTUNE noise for regularization and was trained on a combination of dbbench_sft_dataset_react datasets (v1-v4). Loss was specifically computed only on the assistant's responses, masking user prompts and observations.

Good For

• Developing and evaluating AI agents for complex, multi-step tasks.
• Applications requiring robust tool use and error recovery in simulated environments.
• Research into agentic behavior and multi-turn interaction within LLMs.