open-thoughts/OpenThoughts-Agent-RL-5K
收藏资源简介:
OpenThoughts-Agent-RL-5K是一个包含5,000个任务的强化学习任务集,用于对冷启动监督微调(SFT)模型进行强化学习(RL)微调,以生成最终的代理检查点。与SFT数据集包含完整的(任务,轨迹)对不同,该数据集包含可执行的代理任务(来自pymethods2test-large任务池),策略在沙箱中在线尝试;奖励来自测试验证器,而非存储的教师轨迹。数据集是OpenThoughts-Agent SFT-then-RL配方中的RL任务部分,旨在用于在线策略RL,由RL训练器消费,通过验证器评分并更新策略。它包含5,000行数据,每行对应一个任务,任务标识符为路径字段,任务二进制文件为gzip压缩的任务包(环境定义、源代码和测试),由terminus-2框架在Daytona沙箱中运行。
OpenThoughts-Agent-RL-5K is a 5,000-task reinforcement-learning task set used to RL-finetune the cold-start SFT model into the final agentic checkpoint. Unlike the SFT datasets, which hold full (task, trajectory) pairs, this dataset holds executable agentic tasks (the pymethods2test-large task pool) that the policy attempts on-line inside sandboxes; reward comes from the test verifier, not from a stored teacher trajectory. It is the RL-task half of the OpenThoughts-Agent SFT-then-RL recipe, intended for on-policy RL, consumed by an RL trainer that rolls out the policy against each task, scores it with the verifier, and updates the policy. The dataset contains 5,000 rows, each with a task identifier (path) and a gzip-compressed task bundle (task_binary) that includes environment definition, source, and tests, run under the terminus-2 harness inside Daytona sandboxes.




