主题路线Topic route
数据与训练,按顺序找资料Data and training: resources in order
训练问题很少是「模型不够大」,更常见的是数据授权不清、训练集与测试集串了、污染没查、评测口径不一致。这条路线先处理数据,再谈训练与评测。Training problems are rarely just “the model is too small”; more often the data licence is unclear, train and test overlap, contamination is unchecked, or the evaluation basis differs. This route handles data first, then training and evaluation.
路线顺序不代表推荐排名;没有实测的能力对比、成本降幅与成功率,站群一律写未知。Route order is not a ranking; where no measured comparison, cost reduction or success rate exists, the sites record it as unknown.
先做什么Where to start
按环节找全站资料Find material by stage
边界Boundary
数据集卡只记录官方许可与已知 caveat,不替你判断「你的用途是否合规」;评测卡也不把厂商分数当独立复现。训练成本随硬件、批量、序列长度和利用率变化,本站只给口径与计算器,不给单一报价。Dataset cards record official licences and known caveats; they do not decide whether your use is compliant. Benchmark cards do not treat vendor scores as independent reproduction. Training cost varies with hardware, batch, sequence length and utilisation; the sites give the basis and a calculator, not a single quote.