elliottash/megaman-maker-levels

Mega Man Maker关卡语料库是一个统一的数据集，包含179,166个用户创作的Mega Man Maker关卡，用于训练MarioGPT风格的文本到关卡生成模型、程序内容生成研究以及平台游戏关卡分析。每个数据行包含：MarioGPT风格的ASCII瓦片网格（level_ascii），可直接输入GPT2/BPE分词器；基于关卡结构特征自动生成的的自然语言文本提示（text_prompt）；完整的解析中间表示（IR）作为gzip压缩的JSON（ir_json_gz）；原始.mmlv字节（mmlv_bytes）（如果保留）；完整的webmeka.io互动元数据（如点赞、下载、众包难度等）；以及指向关卡webmeka.io页面的链接（webmeka_url）。数据集提供多个配置子集，包括默认集、受欢迎关卡集、包含Boss的关卡集和与MarioGPT兼容的关卡集。数据来源于webmeka.io上托管的用户创作关卡，仅包含符号ID和元数据，不包含Capcom的资产，并基于CC-BY-4.0许可发布。A unified corpus of 179,166 user-authored Mega Man Maker levels for training MarioGPT-style text-to-level models, procedural-content generation research, and platformer-level analysis. Each row carries: a MarioGPT-style ASCII tile grid (level_ascii) ready to feed directly into a GPT2/BPE tokenizer; a natural-language text_prompt automatically composed from the levels structural features; the full parsed intermediate representation (IR) as gzipped JSON (ir_json_gz); the original raw .mmlv bytes (mmlv_bytes) when retained; the full webmeka.io engagement metadata (likes, downloads, crowd-sourced difficulty, etc.); and a webmeka_url link back to the levels webmeka.io page. The dataset includes config subsets such as default, liked levels, boss levels, and MarioGPT-compatible levels. Source content is user-authored levels hosted on webmeka.io, containing only symbolic IDs and metadata, no Capcom assets, and released under CC-BY-4.0 license for research and PCGML use.