stevenkhan/pokemon-showdown-battle-sft

这是一个包含500,057个来自高ELO（1400+）竞技Pokémon Showdown比赛的专业战斗决策的数据集，用于监督微调语言模型。数据集将原始的Pokémon Showdown对战日志转换为指令跟随对，其中用户消息描述当前战斗状态，助手消息推荐最优行动并附带战略推理。数据集仅包含获胜玩家的决策，模型从成功中学习。数据来源自milkkarten/pokechamp，经过过滤（ELO≥1400、有赢家的游戏、每场游戏≥3个有意义的回合、仅获胜玩家的视角）。数据集覆盖多个世代和格式，包括OU、RU、Ubers、National Dex、Random Battles等。每个示例都是一个聊天消息列表，包含系统、用户和助手角色。战斗状态跟踪包括活跃Pokémon、状态条件、统计提升、已知招式、能力、物品、Mega进化/Terastallization、替补Pokémon、侧面条件、场地条件、世代和格式等。

A dataset containing 500,057 expert battle decisions from high-ELO (1400+) competitive Pokémon Showdown matches, formatted for supervised fine-tuning of language models. The dataset converts raw Pokémon Showdown replay logs into instruction-following pairs where user messages describe the current battle state and assistant messages recommend the optimal action with strategic reasoning. Only the winning players decisions are included — the model learns from success. The source data is processed from milkkarten/pokechamp, filtered by ELO ≥ 1400, games with a winner, ≥3 meaningful turns per game, and winning players perspective only. The dataset covers multiple generations and formats including OU, RU, Ubers, National Dex, Random Battles, etc. Each example is a list of chat messages with system, user, and assistant roles. Battle state tracking includes active Pokémon, status conditions, stat boosts, known moves, abilities, items, Mega Evolution/Terastallization, bench Pokémon, side conditions, field conditions, generation, and format.