Papers12:03Chinese contentPPL-Factory:任务感知与预算感知的数据选择方法想省训练成本又想模型推理更强?PPL-Factory用10%数据在MATH上比全量微调高4.8,门槛低、效果还更炸。#PPL-Factory#数据选择#微调#推理模型aarXiv cs.LG@Hang Zhang, Warren J. GrossSourceRead laterReadWorth followingUsefulFollow PPL-Factory