专属模型全生命周期管理Lifecycle management for dedicated models
数据、评分、训练、报告、上线和迭代,都在同一个控制台里,团队和智能体用同一套规则协作。Data, grading, training, reports, serving and improvement in one console, where teams and agents follow the same rules.

控制台界面The console
四项指标一眼看清Four verdicts at a glance
更好、盲评胜率、更便宜、更快,每一项都有依据可以点开,AI 写好结论和建议。Better, blind win rate, cheaper, faster. Each links to its evidence, and an AI summary states the conclusion.
每次训练都可追溯Every run is traceable
底座、数据快照、评分器、训练栈版本、卡时和花费都记录在案,一条命令复现。Base, data snapshot, graders, stack versions, GPU hours and cost are on record, reproducible with one command.
灰度、影子、回滚Canary, shadow, rollback
别名指向具体版本,回滚只改指向,客户端不用改任何配置。An alias points to a version, so a rollback changes the pointer and clients change nothing.
今天该做什么,一目了然What to do today, at a glance
在用模型、训练实验、达标报告和节省金额,加上下一步建议。Models in use, runs, reports at parity and savings, with the next steps suggested.
上线后与通用大模型统一路由Routed together with general models
路由层按质量、成本和延迟为每次请求挑模型。专属模型拿不准的题,自动回退给通用模型。The router picks a model for every request by quality, cost and latency, and falls back to a general model when the dedicated model is unsure.
$ factory train train-5ed17a9d --purpose train:content-copy \ --base Qwen/Qwen2.5-7B-Instruct --method grpo ✓ 预检通过preflight passed grant, no_leak, graders, licence, budget $ 最坏花费worst case $49.90, 上限cap $100 ● 2×H100 step 30/30 test 0.923 ✓ 已登记模型版本model version registered mv-grpo-20261110 $ factory report build eval-outputs --model-version mv-grpo-20261110 ✓ rpt-20261110-f9cf13 已冻结,达到对等门槛frozen, parity reached
四种接入方式,同一套权限与规则Four entry points, one set of rules
企业的智能体可以自行建数据集、发起训练、读报告;它想上线时只会提交审批,由人来拍板。Agents can build datasets, start training and read reports. When one wants to go live, it files an approval and a person decides.
支持主流开源底座Built on the leading open models
底座按许可证管理,每个底座带好自己的对话模板,训练和推理用同一套。Bases are cleared by licence, and each carries its own chat template, used the same way in training and serving.
关键操作需人工审批Key actions require approval
上线审批Go-live approval
灰度、影子流量和全量都由管理员批准,申请人和批准人都记录在案。Canary, shadow traffic and full rollout are approved by an admin, with both names on record.
花费有上限Spending has a cap
开卡前先算最坏花费,超出预算不启动;机器到时限自动释放。Worst-case cost is checked before any GPU starts, and every machine releases itself at its time limit.
全程可追溯Everything is traceable
谁在什么时候用哪份数据做了什么,逐条可查,随时一键回滚。Who used which data for what, and when, line by line, with one-action rollback.
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.