stackexchange-unix-sandboxes-verified
收藏资源简介:
laion/stackexchange-unix-sandboxes-verified是一个包含10,000个计算机使用任务的数据集,专为Harbor/TaskTrove任务框架设计。它基于Unix & Linux StackExchange的实际问题构建,涵盖shell命令、CLI工具、文件系统和Linux/BSD系统管理等主题。数据集从DCAgent/stackexchange-unix-sandboxes-skywork-response重新打包而来,保留了原始的任务指令文件(instruction.md)和共享的ubuntu:24.04运行环境,但将原有的Skywork奖励模型验证器替换为LLM法官验证器。每个任务包含两列数据:path(字符串类型)和task_binary(gzip压缩包格式)。核心特点是每个任务沙箱都集成了从OpenThoughts-Agent框架移植的LLM法官验证器(verifiers/llm_judge.py)。在验证阶段,系统读取代理生成的/app/response.txt文件以及/tests/verifier_data.json文件(包含任务指令和适配的评分标准),通过litellm调用LLM法官(默认使用openai/gpt-4o-mini模型)进行评估,最终将0.0-1.0的奖励分数写入/logs/verifier/reward.txt。使用该数据集需要配置OPENAI_API_KEY(通过task.toml中的[verifier].env设置传播到验证器容器),并可通过JUDGE_MODEL环境变量覆盖默认的法官模型。评分标准围绕Unix/Linux shell请求的正确性、完整性、安全性和相关性四个维度制定。数据集适用于文本生成、智能体开发、强化学习以及LLM评估等任务场景。
laion/stackexchange-unix-sandboxes-verified is a dataset containing 10,000 computer usage tasks, specifically designed for the Harbor/TaskTrove task framework. It is built based on real-world problems from Unix & Linux StackExchange, covering topics such as shell commands, CLI tools, file systems, and Linux/BSD system management. The dataset is repackaged from DCAgent/stackexchange-unix-sandboxes-skywork-response, retaining the original task instruction files (instruction.md) and the shared ubuntu:24.04 runtime environment, but replacing the original Skywork reward model verifier with an LLM judge verifier. Each task includes two columns of data: path (string type) and task_binary (gzip compressed format). The core feature is that each task sandbox integrates an LLM judge verifier (verifiers/llm_judge.py) ported from the OpenThoughts-Agent framework. During the verification phase, the system reads the agent-generated /app/response.txt file and the /tests/verifier_data.json file (containing task instructions and adapted scoring criteria), then evaluates via litellm calling the LLM judge (default using the openai/gpt-4o-mini model), ultimately writing a reward score of 0.0-1.0 to /logs/verifier/reward.txt. Using this dataset requires configuring OPENAI_API_KEY (propagated to the verifier container via [verifier].env settings in task.toml), and the default judge model can be overridden via the JUDGE_MODEL environment variable. Scoring criteria are formulated around four dimensions: correctness, completeness, safety, and relevance of Unix/Linux shell requests. The dataset is suitable for tasks such as text generation, agent development, reinforcement learning, and LLM evaluation.
数据集概述:laion/stackexchange-unix-sandboxes-verified
该数据集是一个面向文本生成任务(text-generation)的二进制任务集合,专注于Unix/Linux系统管理的计算机使用场景。
基本信息
- 许可协议: Apache-2.0
- 任务类别: 文本生成(text-generation)
- 标签: agent, harbor, reinforcement-learning, llm-judge
- 数据集规模: 包含10,000个任务
数据来源与格式
- 数据源自 Unix & Linux StackExchange 社区的问题,主题涵盖shell脚本、CLI工具、文件系统和Linux/BSD系统管理。
- 数据重新打包自
DCAgent/stackexchange-unix-sandboxes-skywork-response,保留了原始instruction.md和共享的ubuntu:24.04环境。 - 数据格式采用 Harbor / TaskTrove 任务schema,包含两个字段:
path(字符串)task_binary(gzip tar 文件)
验证机制
- 数据集使用LLM评估器(LLM judge)替代了原始的Skywork奖励模型。
- 每个任务的沙箱环境包含一个来自OpenThoughts-Agent
data.nemotron_gym框架的LLM-judge验证器(verifiers/llm_judge.py)。 - 验证流程:
- 在验证时,
tests/test_state.py读取agent生成的/app/response.txt和/tests/verifier_data.json(包含任务指令和评分标准)。 - 通过
litellm调用LLM judge(默认模型为openai/gpt-4o-mini)。 - 将0.0-1.0的奖励值写入
/logs/verifier/reward.txt。
- 在验证时,
使用要求
- 运行验证需要设置环境变量
OPENAI_API_KEY,该变量会通过task.toml中的[verifier].env传播到验证器容器。 - 可选地,可以通过
JUDGE_MODEL环境变量覆盖默认的LLM模型。
评估标准
LLM judge的评分基于以下四个维度:
- 正确性(correctness)
- 完整性(completeness)
- 安全性(safety)
- 相关性(relevance)




