frontiersmith-tasks-graded-v1
收藏资源简介:
FrontierSmith — Graded-Reward Harbor Tasks (v1) 是一个包含10个开放式竞争性编程优化任务的数据集,采用OpenThoughts-Agent/Harbor格式。该数据集专为启发式算法优化设计,任务不提供封闭形式的最优解,而是要求开发者编写C++17启发式代码来解决复杂优化问题。每个任务中,代理需要将解决方案写入`/app/solution.cpp`文件,随后由容器内法官在10个隐藏测试用例上进行评估。评估过程会计算每个用例的连续质量比率(范围[0,1],越高越好),取10个比率的平均值,并与预设的TAU阈值比较,最终转换为二进制奖励(平均值≥TAU时奖励为“1”,否则为“0”)。这种“通过”标准要求提交的解决方案超越确定性基线一定幅度,而非达到理论最优。数据集提供了每个任务的基线比率、TAU阈值和参考启发式平均比率等具体数值。技术实现上,验证完全在单个任务容器内进行,无需Docker Compose或侧车法官,所有测试资产(包括测试用例、检查器和生成器)均位于`tests/`目录下以确保环境一致性。数据集还包含手工编写的启发式C++求解器作为参考解决方案,这些求解器均能通过各自的TAU阈值。数据内容源自MIT许可的FrontierCS/Frontier-CS仓库,对应上游算法问题306–315,经过Harbor转换、容器内法官实现、TAU阈值设定和启发式参考解决方案的添加,最终以MIT许可证发布,需注明原始来源归属。
FrontierSmith — Graded-Reward Harbor Tasks (v1) is a dataset containing 10 open-ended competitive programming optimization tasks, using the OpenThoughts-Agent/Harbor format. This dataset is specifically designed for heuristic algorithm optimization: no closed-form optimal solutions are provided for the tasks, and developers are required to write C++17 heuristic code to solve complex optimization problems. For each task, an AI Agent must write its solution to the `/app/solution.cpp` file, which is then evaluated by an in-container judge across 10 hidden test cases. The evaluation process calculates a continuous quality ratio for each test case (ranging from [0, 1], with higher values indicating better performance), takes the average of the 10 ratios, compares it against a preset TAU threshold, and finally converts the result into a binary reward: "1" if the average is ≥ TAU, and "0" otherwise. This "pass" criterion implies that a submitted solution must exceed a deterministic baseline by a certain margin, rather than reaching the theoretical optimal solution. Each task in the dataset has a dedicated baseline ratio, TAU threshold, and reference heuristic average ratio, with their specific values listed in the accompanying documentation. In terms of technical implementation, validation is conducted entirely within a single task container, eliminating the need for Docker Compose or sidecar judges. All test assets (including test cases, checkers, and generators) are stored in the `tests/` directory to ensure environmental consistency. The dataset also provides handcrafted heuristic C++ solvers as reference solutions, all of which can pass their respective TAU thresholds. The dataset content is sourced from the MIT-licensed FrontierCS/Frontier-CS repository, corresponding to upstream algorithm problems 306–315. After undergoing Harbor format conversion, in-container judge implementation, TAU threshold configuration, and addition of heuristic reference solutions, it is finally released under the MIT license, with attribution to the original source required.
数据集概览:FrontierSmith — Graded-Reward Harbor Tasks (v1)
- 名称:FrontierSmith — Graded-Reward Harbor Tasks (v1)
- 许可证:MIT
- 标签:harbor, agent-tasks, competitive-programming, optimization, graded-reward
核心特性
- 任务类型:10个开放式的竞争性编程优化任务,采用OpenThoughts-Agent / Harbor格式。
- 任务格式:智能体需编写C++17启发式解法(写入
/app/solution.cpp),由容器内裁判在10个隐藏测试用例上评分,并将连续分数通过阈值转换为二值奖励(Harbor奖励)。 - 奖励机制:
- 由于问题为开放式优化,不存在封闭形式的最优解,预期采用启发式方法。
- 上游检查器对每个测试用例输出一个[0,1]的连续质量比率(越高越好,不可行解为0)。
- 容器内裁判计算10个比率的平均值,并与任务特定的阈值(TAU)比较:
reward = "1" iff mean_ratio >= TAU, else "0" - “通过”意味着提交的解法的表现超过了确定性基线的一定幅度,而非找到了最优解。
任务数据(基线比率、阈值与启发式平均比率)
| 任务 | 基线比率 | TAU阈值 | 启发式平均比率 |
|---|---|---|---|
| frontiersmith-1 | 0.10 | 0.18 | 0.374 |
| frontiersmith-2 | 0.00 | 0.07 | 0.265 |
| frontiersmith-3 | 0.50 | 0.54 | 0.650 |
| frontiersmith-4 | 0.50 | 0.54 | 0.653 |
| frontiersmith-5 | 0.00 | 0.12 | 0.403 |
| frontiersmith-6 | 0.50 | 0.51 | 0.551 |
| frontiersmith-7 | 0.20 | 0.23 | 0.289 |
| frontiersmith-8 | 0.50 | 0.62 | 0.904 |
| frontiersmith-9 | 0.50 | 0.51 | 0.570 |
| frontiersmith-10 | 0.00 | 0.20 | 0.699 |
- 基线比率为0.50表示“匹配上游确定性裁判基线”,需要比率大于0.50才能超越。
执行与验证
- 容器架构:验证完全在单个任务容器内部执行,无需Docker Compose、go-judge sidecar或主机绑定挂载,确保与Daytona快照兼容。
- 运行流程:
tests/test.sh调用tests/run_judge.py,编译测试专用检查器,运行每个测试用例,解析比率,计算平均值并与TAU比较。 - 共享环境:所有任务共享一个字节完全相同的
environment/Dockerfile,因此所有10个任务仅产生1个唯一的Daytona快照。
参考解法
- 由于不存在黄金解法,每个任务提供了一个由人工/教师编写的启发式C++求解器(
solution/solve.sh),其平均比率均超过TAU。 - 验证结果:10/10的启发式解法均通过其TAU的Daytona oracle门控,所有任务共享1个唯一快照。
归属与许可
- 问题内容(包括题目描述、testlib生成器/检查器、测试用例)源自MIT许可的上游仓库 FrontierCS/Frontier-CS(对应上游算法问题306–315)。
- 仅此固定种子集的10个问题可再分发,遵循MIT许可并需注明出处。
- Harbor转换、容器内裁判、TAU阈值设定及启发式解法为OpenThoughts-Agent的原创内容。




