GoldenGrapeGentleman1/battle-game-agent-full-sft

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

GoldenGrapeGentleman1/battle-game-agent-full-sft is a 4 billion parameter Qwen3-4B model fine-tuned by GoldenGrapeGentleman1. It is specifically designed to act as an agent for turn-based battle games, processing raw battle protocol lines to output strategic moves. This model excels as a competitive tier agent for games like gen9randombattle, capable of direct inference or serving as a base for battle-oriented GRPO.

Loading preview...

Model Overview

GoldenGrapeGentleman1/battle-game-agent-full-sft is a specialized 4 billion parameter language model, built upon the Qwen3-4B architecture. It has been extensively fine-tuned on approximately 2.3 million turn-based battle game replay logs, enabling it to understand and respond to raw battle protocol lines.

Key Capabilities

  • Game Agent Functionality: Processes battle protocol lines to generate appropriate game actions, such as move … or switch ….
  • Specialized Training: Leverages a large dataset of real game replay logs for highly specific domain expertise.
  • Competitive Performance: Designed to operate as a competitive tier agent, particularly noted for gen9randombattle.

Training Details

The model underwent Full SFT (Supervised Fine-Tuning) with merged weights. Training was conducted on AMD Instinct MI300X hardware, utilizing ROCm and bfloat16 precision. The base model is Qwen/Qwen3-4B.

Use Cases

  • Direct Inference: Can be loaded and used immediately for generating game actions in turn-based battle scenarios.
  • Base for GRPO: Serves as a strong foundation for further reinforcement learning methods like GRPO (Generalized Policy Optimization) in battle-oriented AI development.