Super Model

Super Model

自己的超级模型your own super model

以企业自有数据训练,经冻结测试集与专家盲评证明优于通用大模型后上线,并随真实调用持续迭代。Trained on company data, proven against general models on a frozen test set and blind expert review, then served and improved with real traffic.

  1. 授权Grant数据用途与处理区域已签署Use and processing region signed off
  2. 资料Data8,420 条历史文案,已去重脱敏8,420 past copies, deduplicated and masked
  3. 验收题Acceptance set240 道题封存,训练不可见240 items sealed, never trained on
  4. 评分标准Scoring与专家判断一致率 0.92Agrees with expert judgement at 0.92
  5. 训练Training第 2 版,基于 Qwen2.5-14BVersion 2, on Qwen2.5-14B
  6. 效果验收Acceptance胜出最强通用模型,成本为其 27%Beats the best general model at 27% of its cost
  7. 上线Go-liveent-partner-a-copy 可调用ent-partner-a-copy is callable
已上线Liveent-partner-a-copy
  • 客服工作台Support desk
  • 客户管理CRM
  • 文案工具Copy tool
示例Sample每完成一个环节,宝石亮起一面;七面全亮,模型上线。Each finished stage lights one facet; with all seven lit, the model goes live.
01

训练Train

以企业数据做监督微调与强化学习,多个开源底座并行比较。Fine-tuning and reinforcement learning on company data, with several open bases compared.

02

证明Prove

冻结测试集、独立评分器和专家盲评,生成不可修改的对比报告。Frozen test set, independent grader and blind review, in a report that cannot be edited.

03

上线Serve

注册为一个模型名,灰度、影子、回滚,由路由与通用模型协同调度。One model name, with canary, shadow and rollback, routed alongside general models.

04

迭代Improve

真实调用回流为训练数据,按信号触发再训练,版本持续提升。Real traffic returns as training data; signals trigger retraining and versions keep improving.

一个专属模型的生命周期The life of a dedicated model

从文件被外部模型牵走,到在企业边界内形成自己的模型、与通用模型逐项对比、接入现有系统,再到每一版持续变好。From documents pulled out to an external model, to a model of its own formed inside the company boundary, measured against general models, connected to existing systems and improved version by version.

专属模型生命周期Lifecycle01 / 05

    律所条款识别演示,300 题封存考卷;响应为首字 P95;成本为每千次任务,示例价目。星环为示意。Law-firm clause demo, 300-item sealed exam; latency is P95 first token; cost per 1,000 tasks at sample prices. Rings are illustrative.

    通用大模型缺少企业的术语、格式与口径General models lack the company's terms, formats and voice

    企业的术语、格式和品牌口径,靠提示词每次去教既贵又不稳定;高频任务按旗舰模型计费,成本居高不下。专属模型把这些一次训练进模型里。Teaching a general model a company's terms, formats and voice through prompts is costly and unreliable, and frequent tasks are billed at flagship prices. A dedicated model learns them once.

    为什么要训练专属模型Why train a dedicated model
    任务Task新品:益生菌冻干粉,场景:上新预告New product: freeze-dried probiotics, launch teaser
    通用大模型General model示例Sample

    1.10 s¥38 / 1k
    专属模型Dedicated model示例Sample

    0.42 s¥4.2 / 1k

    质量、成本、延迟与所有权Quality, cost, latency and ownership

    一份对比报告样例中,专属模型质量高于最强通用模型,每千次成本约为其九分之一,首字响应快一倍以上。In a sample report, the dedicated model scored above the strongest general model at about a ninth of the cost per thousand calls, with first tokens more than twice as fast.

    查看全部价值See all benefits
    0.91质量,最强通用模型 0.89Quality, strongest general model 0.89
    1/9每千次成本Cost per 1k calls
    0.42 sP95 首字,对比 1.10 sP95 first token, against 1.10 s
    62%专家盲评胜率Blind expert win rate
    怎么做到How it is done

    平台覆盖专属模型从数据到上线的全过程One platform from data to production

    数据、评分、训练、报告、上线与迭代在同一个控制台完成,每一步都留有记录,任意一次训练都可以一条命令复现。Data, grading, training, reports, serving and improvement in one console. Every step is recorded and any run can be reproduced with one command.

    模型别名Model alias对比报告Comparison report底座 + 适配器Base + adapter合成样本Synthetic samples数据快照Data snapshot
    技术核心Technical core

    SuperLoop:生产在环的企业模型体系SuperLoop: an enterprise model system with production in the loop

    训练信号来自生产环境中的路由裁决:同一请求由多个模型候选作答,业务方的采纳结果被记录下来,同时用作训练奖励、课程来源、评测基准和反作弊审计。模型按周期从这条信号流中更新。Training signal comes from routing verdicts in production: several candidate models answer the same request and the business outcome is recorded. That record serves as reward, curriculum, benchmark and anti-gaming audit, and the model is updated from it on a regular cycle.

    控制面:数据放置判定、评分器隔离、审计Control plane: placement checks, grader separation, audit 客户流量Traffic 生产面Production 路由即实时评测Routing as live evaluation 影子、灰度、全量接管Shadow, canary, full rollout 进化面Evolution 预训练、蒸馏、强化学习、Pre-training, distillation, RL, 偏好对齐,算子可替换alignment; operators swappable 资产面Assets 评分器、冻结评测集、轨迹Graders, frozen sets, traces 裁决、难例、倾向得分Verdicts, hard cases, propensities 候选模型晋升Candidate promotion
    1. RAL
      路由裁决学习Router-arbitrated learning以生产路由中的多模型对比作为组内比较,并用倾向得分纠正路由带来的选择偏差。Multi-model comparisons from production routing act as the group comparison, with propensity scores correcting routing bias.
    2. TAP
      教师锚定后验Teacher-anchored posterior蒸馏与强化学习统一为一个目标:证据充分处偏离教师,证据稀疏处回落到教师。Distillation and RL under one objective: move away from the teacher where evidence is strong, stay close where it is thin.
    3. FFC
      市场失灵课程Frontier-failure curriculum训练预算集中在所有候选模型都未被采纳的请求上。Training budget goes to requests where no candidate model was accepted.
    4. CaV
      合规即验证器Compliance as verifier数据放置、脱敏与出境规则进入训练奖励和报告评分器。Placement, masking and cross-border rules enter the training reward and the report grader.
    5. Gap
      评分器背离审计Hacking-gap audit对比训练评分、封存报告评分与真实采纳胜率三条线,度量评分器被钻空子的程度。Training score, sealed report score and real acceptance rate are compared to measure grader gaming.

    上线后基于真实调用持续迭代Improvement from real traffic after launch

    真实调用与人工修改回流为训练数据;效果漂移、否决率上升或新偏好攒够时触发再训练,新版本先影子对照、再灰度晋升。Real calls and human edits return as training data. Drift, rising rejections or enough new feedback trigger retraining, and new versions run in shadow before a canary.

    了解持续迭代About continuous improvement
    78.1% 92.3%
    测试集准确率Test accuracy训练奖励Train reward

    真实实验:Qwen2.5-7B,GSM8K 测试集,30 步强化学习,两张 H100,约 8 美元。Real run: Qwen2.5-7B, GSM8K test set, 30 RL steps, two H100s, about $8.

    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.