tulu3-sft-personas-math-sandboxes-verified
收藏资源简介:
该数据集名为laion/tulu3-sft-personas-math-sandboxes-verified,是一个用于文本生成任务的数学应用题数据集,特别适用于智能体(agent)、Harbor平台、强化学习和数学相关的研究。数据集包含9,998个任务,每个任务由两列数据构成:path(字符串类型)和task_binary(gzip压缩的tar包格式),遵循标准的Harbor/TaskTrove任务模式。数据来源于Tulu-3 SFT personas-math,是DCAgent/tulu3-sft-personas-math-sandboxes的镜像,内容保持不变。每个任务都配备了原始的确定性验证器(tests/test.sh),用于评估智能体生成的答案:智能体将最终答案写入/app/answer.txt,验证器通过精确字符串匹配(去除空格)将其与每个任务的标准答案进行比较,结果映射为奖励1(正确)或0(错误)。该验证过程完全基于确定性规则,无需LLM评判,因此不产生API成本。所有任务在共享的python/ubuntu环境中运行。
The dataset named laion/tulu3-sft-personas-math-sandboxes-verified is a mathematical word problem dataset for text generation tasks, particularly suitable for agent, Harbor platform, reinforcement learning, and mathematics-related research. It contains 9,998 tasks, each consisting of two columns: path (string type) and task_binary (gzip-compressed tar package format), following the standard Harbor/TaskTrove task pattern. The data is sourced from Tulu-3 SFT personas-math and is a mirror of DCAgent/tulu3-sft-personas-math-sandboxes, with content remaining unchanged. Each task is equipped with an original deterministic verifier (tests/test.sh) to evaluate answers generated by agents: the agent writes the final answer to /app/answer.txt, and the verifier compares it with the standard answer for each task via exact string matching (with spaces removed), mapping the result to a reward of 1 (correct) or 0 (incorrect). This verification process is entirely rule-based, requiring no LLM judgment and thus incurring no API costs. All tasks run in a shared python/ubuntu environment.
数据集概述
数据集名称:laion/tulu3-sft-personas-math-sandboxes-verified
许可证:Apache-2.0
任务类别:文本生成
标签:agent, harbor, reinforcement-learning, math
数据集规模:9,998 个任务
数据格式:包含两列——path(字符串)和 task_binary(gzip tar 格式),遵循规范的 Harbor / TaskTrove 任务模式。
数据来源:源自 Tulu-3 SFT personas-math 的数学应用题任务。此数据集是 DCAgent/tulu3-sft-personas-math-sandboxes 的镜像副本,内容未作修改。
评估机制:每个任务附带原始的确定性验证器(tests/test.sh)。智能体需将最终答案写入 /app/answer.txt,验证器通过精确字符串匹配(去除空白字符)与任务的黄金答案进行比较,输出奖励值为 1(正确)或 0(错误)。该过程不涉及大语言模型评判,完全确定,无需 API 成本。
运行环境:共享的 python/ubuntu 环境。




