DCAgent/code-contests-noblock
收藏资源简介:
code-contests-noblock是一个基于CodeContests衍生的竞争性编程强化学习任务数据集,包含8,728个任务,以Harbor格式存储。每个任务以gzip压缩的tar文件形式保存在tasks.parquet文件中,包含路径(path,格式为<task_id>.tar.gz)和任务二进制数据(task_binary)两列。每个tar文件包含以下组件:instruction.md(任务说明)、task.toml(任务配置)、environment/Dockerfile(环境Docker文件)以及一个tests/验证器目录,其中包含test.sh(测试脚本)、test_state.py(测试状态代码)和test_data.json(测试数据)。该数据集专为文本生成任务设计,适用于代理(agent)、代码生成、代码竞赛、Harbor平台、强化学习(RL)和RL验证器(rlvr)等应用场景。v2版本修复了验证器奖励记录问题,确保在任务测试失败时也能正确记录奖励值(0表示失败,1表示成功),避免了v1版本中因脚本提前终止导致的奖励分布偏差。
code-contests-noblock is a competitive programming reinforcement learning (RL) task dataset derived from CodeContests, which includes 8,728 tasks and is stored in Harbor format. Each task is saved as a gzip-compressed tar file, and the tasks.parquet file contains two columns: "path" (formatted as <task_id>.tar.gz, which records the storage path of each task's tar archive) and "task_binary" (which stores the task's binary data). Each tar archive contains the following components: instruction.md (task description), task.toml (task configuration), environment/Dockerfile (environment Dockerfile), and a tests/ validator directory that includes test.sh (test script), test_state.py (test state code), and test_data.json (test data). This dataset is specifically designed for text generation tasks, and is applicable to scenarios including AI agents, code generation, competitive programming contests, Harbor platform, reinforcement learning (RL), and RL validator (rlvr). Version 2 fixes the validator reward recording issue, ensuring that reward values can be correctly recorded even when task tests fail, where 0 indicates failure and 1 indicates success, thus avoiding the reward distribution bias caused by early script termination in Version 1.




