mind-of-tashi-runs
收藏资源简介:
The Mind of Tashi — leaderboard runs 数据集是一个持久化、公开、仅追加的运行日志,专门用于支持HuggingFace Space中The Mind of Tashi游戏的排行榜功能。它记录了玩家在游戏中的完整运行数据,每个已完成的运行对应一个JSON文件,存储在runs/目录下,采用<unix_ts>-<run_id>.json的命名格式,以确保并发提交时不会相互覆盖。数据集包含详细的运行统计信息,包括:运行ID(12位十六进制UUID)、时间戳(Unix秒)、用户名(最多40字符)、来源(已验证的hf-oauth或访客guest)、总回合数、总耗时(秒)、每层游戏数据(对手ID、回合数、耗时、是否获胜)、是否通关完整阶梯,以及对手使用的后端(llama.cpp或mock)。数据集规模小于1万样本,采用CC BY-SA 4.0开源许可证。身份验证机制确保数据真实性:在部署的Space中,OAuth会话用户名会覆盖客户端提交的内容,防止冒名提交;本地运行时则信任输入的用户名。排行榜排名规则仅考虑获胜的运行,保留每个用户的最佳通关记录(优先按最少总回合数排序,平局时按最短耗时排序),最多显示20条记录。所有尝试记录(包括失败)均被保存以支持分析,但只有获胜记录参与排名。该数据集是Build Small Hackathon, Track Two项目提交的六个组件之一,与游戏Space、自对弈数据集、SFT模型+GGUF、OpenEnv强化学习环境、GRPO模型+GGUF等其他五个组件配套使用,共同构成完整的游戏AI开发与评估体系。
The Mind of Tashi — leaderboard runs dataset is a persistent, public, append-only log of runs specifically designed to support the leaderboard functionality for the game The Mind of Tashi on HuggingFace Space. It records complete run data from players, with each completed run corresponding to a JSON file stored in the runs/ directory under the naming format <unix_ts>-<run_id>.json, ensuring no overlap during concurrent submissions. The dataset includes detailed run statistics: run ID (12-digit hexadecimal UUID), timestamp (Unix seconds), username (up to 40 characters), source (authenticated hf-oauth or guest guest), total rounds, total time taken (seconds), per-level game data (opponent ID, rounds, time taken, win/loss), whether the full ladder was completed, and the backend used by opponents (llama.cpp or mock). The dataset has fewer than 10,000 samples and is licensed under CC BY-SA 4.0. An authentication mechanism ensures data authenticity: in deployed Spaces, OAuth session usernames override client-submitted content to prevent spoofing, while local runtime trusts input usernames. Leaderboard ranking rules only consider winning runs, retaining the best completion record per user (prioritized by fewest total rounds, tied by shortest time taken), with a maximum of 20 records displayed. All attempts (including failures) are saved for analysis, but only winning records participate in ranking. This dataset is one of six components submitted for the Build Small Hackathon, Track Two project, used alongside other components such as the game Space, self-play dataset, SFT model+GGUF, OpenEnv reinforcement learning environment, and GRPO model+GGUF, forming a complete game AI development and evaluation system.
数据集概述
数据集名称:The Mind of Tashi — leaderboard runs
许可证:cc-by-sa-4.0
数据集规模:少于10,000条记录(n<10K)
数据集标签:leaderboard, game
数据集描述
该数据集是“The Mind of Tashi”游戏排行榜的持久化、公开、仅追加的运行日志。每个完成的游戏运行对应一个JSON文件,存储在 runs/ 目录下,确保并发用户不会相互覆盖提交。
数据布局
runs/ <unix_ts>-<run_id>.json
每个文件名由Unix时间戳和运行ID组成,代表一次完整的游戏记录。
数据字段(Run schema)
| 字段 | 类型 | 说明 |
|---|---|---|
run_id |
str | 12位十六进制UUID |
ts |
int | Unix时间戳(秒) |
username |
str | 用户名,最多40个字符 |
source |
str | hf-oauth(已验证)或 guest(手动输入) |
total_turns |
int | 整个运行的总回合数 |
total_seconds |
float | 实际耗时(墙钟时间) |
per_level |
list | 每一级的对手信息:{opponent_id, turns, seconds, won} |
won |
bool | 是否通关整个天梯 |
backend |
str | 使用的推理后端:llama.cpp 或 mock |
身份验证策略
- 在已部署的Space中,OAuth用户名会覆盖客户端提交的用户名,
source标记为hf-oauth。 - 本地运行时(未设置
SPACE_ID),手动输入的用户名被信任,source标记为guest。 - 排行榜界面会在已验证的行上显示 ✓ hf 徽章。
排行榜排名规则
- 仅统计
won为true的运行。 - 每个用户只保留最佳通关记录(
total_turns最少,若平局则按total_seconds最低)。 - 最多返回20条记录。
- 所有尝试(包括失败)均被存储以供分析,但只有获胜记录参与排名。
所属项目
该数据集是 Build Small Hackathon, Track Two 的六个工件之一,配套资源包括:
- 游戏Space:
build-small-hackathon/mind-of-tashi-* - 自对弈数据集
- 微调模型及GGUF
- OpenEnv Gym环境
- GRPO模型及GGUF





