Overview

The thetmon/c23 is a LoRA adapter for the Qwen/Qwen3-4B-Instruct-2507 base model, developed by thetmon. This adapter, with 4 billion parameters, is specifically fine-tuned to improve multi-turn agent task performance across diverse domains.

Key Capabilities

• Enhanced Agentic Reasoning: Optimized for complex, multi-step tasks requiring sequential decision-making.
• Multi-turn Task Proficiency: Excels in scenarios where the agent needs to interact with an environment over multiple turns, such as household tasks (ALFWorld) and database operations (DBBench).
• Error Recovery: Training includes loss application to all assistant turns, enabling the model to learn from and recover from errors within a trajectory.
• Tool Use Integration: Designed to facilitate effective tool use and action selection based on environmental observations.

Training Details

The adapter was trained using LoRA (r=64, alpha=128) on a full precision base model, with a maximum sequence length of 4096 tokens over 3 epochs. The training data combined u-10bei/sft_alfworld_trajectory_dataset_v5 and u-10bei/dbbench_sft_dataset_react_v4.

Good For

• Developing AI agents that require robust multi-turn interaction.
• Applications involving complex task automation in simulated or real-world environments.
• Scenarios demanding improved tool use and error handling in agentic workflows.