deepkick/qwen3-4b-advanced-sft-v13-merged

Overview

This model, deepkick/qwen3-4b-advanced-sft-v13-merged, is a 4 billion parameter language model derived from the Qwen/Qwen3-4B-Instruct-2507 base model. It has undergone a LoRA SFT (Supervised Fine-Tuning) process, with the adapter merged directly into the base model for seamless deployment.

Key Capabilities

• Advanced Agentic Task Performance: Specifically fine-tuned using the u-10bei/sft_alfworld_trajectory_dataset_v5, which focuses on complex trajectory-based tasks, making it suitable for agentic applications.
• Optimized for AgentBench: Designed with a particular focus on performance within AgentBench Advanced evaluations, indicating its strength in environments requiring sophisticated decision-making and planning.
• vLLM Compatibility: The LoRA adapter has been merged, and there are no tokenizer vocabulary modifications, ensuring compatibility with vLLM for efficient inference.

Training Details

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Dataset: u-10bei/sft_alfworld_trajectory_dataset_v5
• Method: LoRA SFT (merged)
• Max Sequence Length: 4096 tokens
• Epochs: 1
• Learning Rate: 1e-06
• LoRA Configuration: r=32, alpha=128

Good For

• Developers working on AI agents requiring robust performance in complex, trajectory-based environments.
• Researchers and practitioners involved in AgentBench Advanced evaluations.
• Applications demanding a Qwen3-4B variant with enhanced capabilities for structured, sequential reasoning.