Overview

thetmon/c10 is a LoRA adapter (r=64, alpha=128) specifically fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model. Developed by thetmon, this 4 billion parameter adapter focuses on enhancing the base model's capabilities for complex, multi-turn agent tasks. It leverages LoRA with Unsloth for efficient training.

Key Capabilities

• Multi-turn Agent Performance: Significantly improves the model's ability to handle sequential, interactive tasks.
• Task Specialization: Optimized for household tasks (ALFWorld) and database operations (DBBench).
• Learning Trajectories: Trained to learn from all assistant turns in a multi-turn trajectory, covering environment observation, action selection, tool use, and error recovery.

Good for

• Developing AI agents that require robust multi-turn interaction.
• Applications involving complex task execution in simulated environments.
• Scenarios where models need to learn from and adapt to environmental feedback over multiple steps.

This adapter provides specialized performance for agentic workflows by focusing on the nuances of multi-turn interactions, making it a strong candidate for developing intelligent agents.