Model Overview

This model, developed by Mountaingorillas, is a fully merged fine-tune of Qwen/Qwen2.5-7B-Instruct, featuring 7.6 billion parameters and a 32K context window. Unlike adapter-only versions, it can be loaded directly. Its core innovation lies in its optimization for multi-turn agent tasks, particularly for the LLM2025 Agent competition, targeting ALFWorld (household tasks) and DBBench (database operations).

Key Capabilities & Innovations

• Hybrid Reasoning Schema (Data Mixing): The model is trained to dynamically adapt its inference format based on the prompt context, preventing common agentic failure modes like parsing errors.
• DBBench Optimization: Strictly adheres to the ReAct format for database operations, ensuring precise SQL string syntax.
• ALFWorld Optimization: Employs native Function Calling (tool_calls for the act function) for robust environment interactions, avoiding invalid action errors.
• Multi-turn Learning: Loss is applied across all assistant turns in a trajectory, enhancing its ability to learn observation, action selection, tool use, and error recovery.

Training Details

• Base Model: unsloth/Qwen2.5-7B-Instruct
• Method: LoRA (merged into full weights)
• Epochs: 3
• Max Sequence Length: 3072

Good For

• Developing and testing agents for complex, multi-turn tasks.
• Applications requiring precise adherence to structured output formats (e.g., SQL generation, function calls).
• Research into agentic AI and hybrid reasoning strategies.