GoldenGrapeGentleman1/pokemon-showdown-agent-full-sft

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

GoldenGrapeGentleman1/pokemon-showdown-agent-full-sft is a 4 billion parameter Qwen3-based language model fine-tuned on 2.3 million Pokémon Showdown replay logs. This model is specifically designed to read raw battle protocol lines and output the next action, such as 'move' or 'switch'. It excels as a competitive tier agent for gen9randombattle, providing direct inference for battle actions.

Loading preview...

Overview

This model, GoldenGrapeGentleman1/pokemon-showdown-agent-full-sft, is a specialized 4 billion parameter language model built upon the Qwen3 architecture. It has been extensively fine-tuned using 2.3 million Pokémon Showdown replay logs, enabling it to understand and respond to competitive Pokémon battles.

Key Capabilities

  • Battle Protocol Interpretation: Reads raw Pokémon Showdown battle protocol lines.
  • Action Generation: Outputs precise battle actions, such as move … or switch ….
  • Competitive Play: Functions as an agent for the gen9randombattle competitive tier.
  • Base for GRPO: Can be used directly for inference or as a foundational model for battle-oriented GRPO (Generative Reinforcement Learning with Policy Optimization).

Training Details

The model was trained using a full Supervised Fine-Tuning (SFT) method, merging weights from the base Qwen3-4B model. The training utilized an AMD Instinct MI300X with ROCm and bfloat16 precision, processing a large dataset of raw Showdown replay logs.

Good For

  • Developing AI agents for competitive Pokémon Showdown.
  • Researching AI strategies in turn-based competitive gaming.
  • Integrating into larger systems requiring automated Pokémon battle decision-making.

Finetunes
1 models