dataset/train 的四个来源统一转换成 Harbor 任务包,跨来源连续编号。
由 ../convert_train_to_harbor.py 生成,请勿手工编辑单个任务;改转换脚本后重新生成。
python3 ../convert_train_to_harbor.py --force
task_00001 – task_00021
sciprobench
21
RewardKit LLM judge,5 维 rubric
加权 rubric ACC
4 个有参考结果
task_00022 – task_00057
env_long_horizon
36
RewardKit agent judge(claude-code 读工作区)
按 points 加权 rubric ACC
无
task_00058 – task_00092
ws_hardest
35
程序化表格比对 + 单元格级… See the full description on the dataset page:
https://huggingface.co/datasets/aoooa/pj-Odyssey-Arena.