怎么做到How it works

从企业数据到可调用模型的五个步骤Five steps from company data to a callable model

每一步都留记录,任何一次训练都能一条命令复现。往下滚动,看每一步在做什么。Every step leaves a record and any run can be reproduced with one command. Scroll to watch each step.

体检数据Check the data

导入表格或调用记录,去重、脱敏,切出并冻结测试集,算出能不能训、值不值得训。Import spreadsheets or call logs, de-duplicate, mask personal details, freeze a test set, and find out whether training is worth it.

交付:数据体检报告Delivers: a data check-up

合成样本Synthesize samples

数据不够时,许可清楚的教师模型为每道题写多个候选,评分器挑出好的,难题单独成池。When data is thin, licensed teacher models write several candidates per prompt, graders keep the good ones, and hard cases are pooled apart.

专家抽检合格率不低于 90%At least 90% pass expert review

训练模型Train the model

监督微调打底,强化学习冲刺。多个底座并行训练,在验证集上挑出最好的。Supervised fine-tuning first, reinforcement learning to push further. Several bases train in parallel and the best one is picked on validation data.

训练栈版本固定,一条命令复现Pinned stack, one-command reruns

证明效果Prove it

冻结测试集、独立的报告评分器、业务专家盲评。三样齐了才出报告,报告冻结后不可修改。Frozen test set, an independent report grader and blind expert review. All three, or no report, and the report cannot be edited afterwards.

交付:冻结的对比报告Delivers: a frozen comparison report

上线路由Serve and route

注册为 MaaS 里的一个模型名,按比例灰度,指标稳定后全量,随时一键回滚。Registered as a model name in MaaS, rolled out by percentage, promoted when metrics hold, rolled back in one action.

应用只需要换一个模型名Business app changes one model name

证明环节的做法How the proof step works

测试集先冻结The test set is frozen first

训练开始前切出并记录哈希。训练数据逐条比对,改写过的近似题也会被查出来,发现泄漏就不开训。Cut and hashed before training. Training data is checked against it, near-copies included, and a leak stops the job.

教的和考的分开The coach never grades the exam

训练用过的评分器进不了报告。模型评委先和专家校准,一致率不到八成不能用。A grader used in training cannot score the report. Model judges must agree with the company's experts at least 80% of the time.

专家盲评Experts judge blind

隐去模型名,打乱顺序和左右位置,胜率附带置信区间,不只给一个数。Names hidden, order and sides shuffled, and the win rate comes with a confidence interval.

建议从一个高频任务开始评估Start the evaluation with one frequent task

评估从数据体检开始:判断该任务能否训练、是否值得训练,并给出预期成本与回本周期。The evaluation starts with a data check-up: whether the task can be trained, whether it is worth it, and the expected cost and payback.