satoyutaka/Qwen2.5-7B-AgentBench-llm2025_advance_v3-BF16

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Feb 24, 2026Architecture:Transformer Cold

satoyutaka/Qwen2.5-7B-AgentBench-llm2025_advance_v3-BF16 is a 7.6 billion parameter agent model developed by satoyutaka, based on the Qwen2.5-7B-Instruct architecture. It is specifically optimized for the AgentBench-comp competition, focusing on strict format adherence for SQL and ALFWorld tasks and enhanced ReAct loop reasoning. This model excels in complex aggregation, multi-step planning, and error recovery, making it suitable for agentic applications requiring precise command execution and environmental reasoning.

Loading preview...

Overview

satoyutaka/Qwen2.5-7B-AgentBench-llm2025_advance_v3-BF16 is a specialized 7.6 billion parameter agent model, built upon the Qwen2.5-7B-Instruct architecture. Developed by satoyutaka, this V3 model is highly optimized for the AgentBench-comp competition, emphasizing robust reasoning and strict adherence to operational protocols.

Key Capabilities & Features

  • 7B Scale Reasoning: Leverages its larger parameter count for deeper logical analysis and improved understanding.
  • Strict Format Adherence: Fine-tuned to precisely follow Action: Operation for SQL and THOUGHT/ACTION for ALFWorld, crucial for competition evaluation.
  • Enhanced ReAct Loop: Prioritizes "Think before you Act" to ensure agents reason about the environment before executing commands.
  • Standard BF16 Compatibility: Merged into a standard BF16 format for 100% compatibility with NVIDIA-based inference engines like vLLM.
  • Synthetic Training Data: Trained exclusively on 100% synthetic SQL and ALFWorld datasets, focusing on complex aggregation, multi-step planning, and error recovery trajectories.

Development & Compatibility

The model's development involved a complex workflow to ensure compatibility across different platforms, including Mac-efficient training and final merging into a standard BF16 format. A critical note for vLLM compatibility is the manual patching of tokenizer_config.json to remove extra_special_tokens to prevent AttributeError during vLLM server startup.

Good For

  • Agentic applications requiring precise command execution.
  • Tasks involving complex logical analysis and multi-step planning.
  • Environments where strict output format adherence is critical.
  • Research and development in agent-based AI systems, particularly for competition-oriented tasks like AgentBench-comp.

Finetunes
1 models