immortal3/wordmaze
收藏资源简介:
Wordmaze是一个合成单词梯子谜题数据集,专为强化学习实验(特别是小语言模型上的GRPO)设计,旨在提供一种一次尝试、易于评分且需要较多规划的任务。每个数据行包含一个谜题及其已知解决方案路径,无需人工标注即可用于训练或评估。谜题规则包括:从起始词开始,每次改变恰好一个字母,所有单词必须是给定长度的真实英语单词,在恰好max_moves步内到达目标词,并匹配密码模式(密码由字母变化方向F(向前)或B(向后)组成)。数据集包含两个配置:m3-4(单词长度3-4,步数3-4)和m4-6(单词长度4-6,步数4-6),每个配置有2,000个示例,分割为训练集(1,600)、验证集(200)和测试集(200)。字段包括id、start、goal、word_length、max_moves、password、prompt、messages、solution、answer、path和num_moves。数据集基于wordfreq英语前10,000个词条生成,排除专有名词,通过随机游走在单词图中采样路径。适用于约束遵循基准测试、基于验证器的强化学习训练以及符号约束跟踪研究。注意:词典有限,密码约束可能对应多个解,难度随单词长度增加,且内容为结构化提示而非自然语言故事。
Wordmaze is a synthetic word-ladder puzzle dataset with an extra password constraint, built for RL experiments (GRPO on small LMs) to offer a one shot, easy to grade task that requires more planning. Each row contains a puzzle and a known solution path, enabling training or evaluation without human labels. The puzzle involves changing exactly one letter per step, using real English words of a given length, reaching the goal in exactly max_moves steps, and matching a password pattern (composed of F for forward or B for backward letter changes). The dataset includes two configurations: m3-4 (word lengths 3-4, moves 3-4) and m4-6 (word lengths 4-6, moves 4-6), each with 2,000 examples split into train (1,600), validation (200), and test (200). Fields include id, start, goal, word_length, max_moves, password, prompt, messages, solution, answer, path, and num_moves. It is generated from wordfreqs top 10,000 English lemmas (excluding proper nouns) by sampling random walks on a word graph. Suitable for benchmarking constraint-following, verifier-based RL training, and studying symbolic constraint tracking. Limitations: limited dictionary, non-unique solutions, variable difficulty by word length, and structured prompts only (no natural language).





