Jarrodbarnes/processrl-terminal-environments
收藏资源简介:
ProcessRL Terminal Environments(ProcessRL终端环境)是一个用于训练和评估代理进程控制的行为条件终端环境集合。该数据集专注于交互式终端工作中出现的故障,例如在误导性成功命令后停止、重复无效操作、在死胡同后未能转向、丢失迁移状态跟踪以及未完成部分进展。数据集包含可执行环境的源定义,而非预构建的运行时镜像,包括53个训练任务和14个保留任务,每个任务包包含指令文件、元数据、环境设置Dockerfile、最终状态验证器等文件。任务通过分阶段管道生成,涉及行为跟踪、过程信号转换、环境生成、可执行准入、校准和参考过滤。数据集覆盖六个进程控制行为轴,如错误退出代码处理、循环探索、死胡同后重复循环、部分进展停滞、新旧状态混淆和稀疏反馈过早停止。推荐使用任务指令作为策略可见提示,并使用最终验证器进行奖励或评估。数据集是用于可执行训练语料库,而非广泛的通用终端代理基准,环境为合成内容,文件、凭证等均为虚构。
ProcessRL Terminal Environments is a collection of behavior-conditioned terminal environments for training and evaluating agent process control. The tasks are designed around failures that appear in interactive terminal work: stopping after a misleading successful command, repeating an unproductive action, failing to pivot after a dead end, losing track of migrated state, and leaving partial progress unfinished. This release contains the first public train/heldout ProcessRL split, with 53 training tasks and 14 heldout tasks. It includes executable task packages with instructions, container build recipes, and final-state verifiers, but not prebuilt runtime images. The dataset is constructed through a staged pipeline involving behavior traces, process signals, environment generation, executable admission, calibration, and reference filtering. It covers six process-control axes: exit_code_false_success, wander_loop_without_convergence, repeat_loop_after_dead_end, partial_progress_stall, old_state_new_state_confusion, and premature_stop_on_sparse_feedback. Recommended use involves using the task instruction as the policy-visible prompt and the final verifier for reward or evaluation. This release is an executable training corpus focused on process-control behaviors, with synthetic environments containing fictional content.




