中文精读
小模型训练便宜之后：逯雨鑫把难点放回数据与服务
42章经单期精读 · 个人后训练与应用成本
嘉宾：逯雨鑫，Mind Lab研究员（当期介绍）｜主持：曲凯
发布：2026年8月29日｜时长：47分29秒｜整理：2026年10月18日
官方单集与节目介绍
先解释“登顶”和“数百美元”
官方说的登顶，是Hugging Face Model Trending热度榜，不是综合性能第一。逯雨鑫使用现有底座做后训练，而非从零预训练通用模型；某项任务分数提高，也不代表所有能力增强。
他自述一张RTX 5090、参数高效微调方式QLoRA和约两百美元订阅辅助生成数据，真实训练十几个小时，估计租用硬件成本不超过一百美元。第一代约五天，第二代两三周。这里没有计入个人劳动、显卡拥有成本、失败试验和持续服务，不能从标题推导任何团队的总预算。
1. 流程从目标开始，榜单只是其中一项检查
线索：04:52–11:24。 逯雨鑫说某Agent评测从约15提升至55到60，同时承认其他评测有下降。Agent指能使用工具执行任务的模型系统；具体榜单、尺度与测试条件未独立核验，不能译成总体智能提高几倍。
他的流程是：明确任务、选底座、做数据管线、选择训练方法、训练、看错例、再改数据。SFT即监督微调，按给定样例训练模型行为；它在个人预算里较实际，不代表对所有目标都胜过强化学习。
底座还要看语言、架构和部署工具。他遇到某底座中文表现差，是版本与任务经验，不能扩成整个模型家族永久不适合中文。AI编程能协助改脚本，脚本跑通也不能证明数据无泄漏、评测正确或产物可上线。
编辑推论：先写需要改变的用户结果，再选技术。若现有模型加提示、工具或检索已够用，训练不一定是最先该做的工作。榜单改进应回到未参与训练的真实样例验证。
2. 数据最花时间，因为病灶不是一句口号
线索：09:17–15:34。 逯雨鑫估计约95%时间用于数据，第二代约十二天做数据。实际使用真实轨迹约六七成，其余靶向合成，筛完剩数千条。这是项目做法，不是通用最佳配比。
最具体的病灶是“任务没完成却宣告完成”。让生成器简单制作“治疗过度自信”的数据，可能只得到听起来谦虚的句子，没有纠正工具状态与结束判断。需要人工抽查，指出哪里不对，再修改生成约束。
他以五千条中亲审约五百条说明劳动，不能当统计充分性的保证。真实使用、公开材料和社区贡献也不天然拥有同样授权，模型输出与用户日志各有使用范围。
编辑推论：按来源、任务、风险与失败类型分层检查。保留工具回执、最终产物和完成声明的对应关系，比只训练一句“可能没做好”更接近目标。合成数据不能替代真实结果，真实数据也不能跳过隐私与许可。
3. 老师轨迹正确，不代表学生学得会
线索：15:34–19:56。 他用Capacity Gap解释难点：强老师可能调用隐含知识、长推理链或复杂工具，小模型即使模仿文字，也未形成稳定能力。训练前看起来好的数据，训练后仍可能无效。
节目对教师参数的猜测不可靠，本篇不补成架构事实。没有提高也可能来自格式、学习设置、配比或评测问题，不能一概归因学生太小。
逯雨鑫强调按任务目标判断蒸馏好坏：若想改善安全边界，却只刷榜单衍生题，涨分也偏离目标。反过来，目标提高不能成为忽略其他能力退化的理由。转写的on-policy等术语混乱，也不宜照录成精确分类教程。
编辑推论：保持数据、底座、训练设置和评估版本，逐次改变一处可解释因素。既测目标任务，也留语言、通用能力与安全回归集；老师更强不是全部答案，学生实际学到什么才是验收。
4. 训练账和Serving账，必须分开算
线索：20:01–26:36。 主持人追问后续调用，逯雨鑫立即区分服务阶段：部署框架、模型架构、并发和运行环境有更多问题。节目用H200服务十来位员工作粗略例子，未说明购租、计费期、上下文和吞吐，不能当采购配置。
他做小说模型的经历又强调工具链成熟度：当时认为Qwen相关训练与部署坑少，非主流架构更麻烦。这个经验可能影响选择，但不是当前排行；遇到问题应查实际成功做法，不能只相信AI建议。
完整训练平台有机会把脚本和服务接起来，减少重复工作；一键训练得到的仍是候选模型。平台无法代替数据权利、任务定义和业务验收。
编辑推论：总预算应含人力、数据、迭代、部署、峰值、闲置、监控与升级。低调用或波动大时，自部署未必便宜；稳定窄任务与敏感数据也许有不同经济性，但要用实际负载比较。
5. 应用成为“小实验室”，需要可学习的数据闭环
线索：23:44–36:50。 双方设想，应用公司已有用户、场景和经营闭环，可以定向后训练；前沿实验室继续承担昂贵预训练与研究，小团队买服务。这是未来分工假说，不是所有应用都应立即建立三四人训练团队。
有用户记录不等于有监督信号。闲聊、失败点击和难解释偏好，仍需变成任务、结果与可验证反馈。数据越多越值钱，也不是无需筛选的规律。
他们保留开源供给变化的担忧，同时主持人认为竞争会推动开源继续。不能把这种担忧写成开源即将停止，更不能据此要求所有公司仓促训练。
编辑推论：应用能积累的优势，可能是发现真实失败、获得合法数据、验证改进再服务用户。基础模型与应用价值可以一起增长，不必先证明谁会吃掉谁，才开始做小范围验证。
6. Local AI有多种价值，也有不同代价
线索：36:59–46:24。 本地模型可能服务敏感数据与高消耗创作，减少向模型供应商发送信息。逯雨鑫也承认当时自己本地使用率不高，错误与能力仍有差距；制作人追问最前沿能力，曲凯追问普通人是否真会为隐私迁移，最后把便利和应用默认选择也纳入讨论。
本地不是“永不外流”：遥测、插件、日志、同步和工具仍可能发送数据。设备已经买下，也不代表运行、维护和用电成本为零。片中未来手机模型、苹果内存及匿名泄漏案例均未核，不作为产品承诺或指控。
编辑推论：把隐私、离线可用、质量、延迟、费用与使用便利分别测。敏感业务可能有更强部署理由，大众可能更在意顺手；两类选择可以并存，不能由一类用户替所有人决定。
给一次微调写完整成本与验收记录
下面是编辑练习。选一个稳定窄任务，固定独立评估和旧任务回归，比较现有API、提示与工具调整、检索、微调几条路径。保留错误分类、完成证据、数据授权和版本，再记录数据劳动、训练迭代与服务费用。
若目标提高但复核更重、旧能力退化或总成本上升，就继续定位原因，不用热度榜替代结果。只有离线验证通过，才讨论实际部署。这是编辑方法，不是本篇已复现的训练教程。
回听导航与关联阅读
- 03:23：个人项目的成本范围。
- 06:08：从目标到训练的流程。
- 11:33：数据与“过度自信”病灶。
- 15:34：容量差距。
- 20:01：训练和服务分账。
- 27:01：应用实验室的分工假说。
- 40:17：前沿能力、本地与便利的争论。
相关主题：小模型、数据迭代、完整成本。可对照孟繁青谈数据质量和朱亦博谈真实运行成本，检查训练成绩怎样进入服务结果。
资料边界：完整阅读所供转写、旧精读与官方介绍，未逐句复听、读模型卡、查计费或执行训练。热度榜与性能分开，成本与分数为嘉宾自述；含混底座、教师版本、部署配置与未来硬件不猜补，不构成复现保证或采购建议。
English edition
After Small-Model Training Gets Cheaper: Lu Yuxin Returns to Data and Serving
42 Zhangjing episode analysis · Individual post-training and application costs
Guest: 逯雨鑫, researcher at Mind Lab, as introduced at release | Host: Qu Kai
Published: August 29, 2026 | Duration: 47 minutes 29 seconds | Prepared: October 18, 2026
Official episode and introduction
Explain “number one” and “hundreds of dollars” first
The official claim concerns Hugging Face’s Model Trending popularity list, not first place in overall performance. Lu post-trained an existing model rather than pretraining a general model from scratch. Improvement on a task does not establish stronger capability everywhere.
Lu reports one RTX 5090, parameter-efficient fine-tuning through QLoRA and a roughly US$200 subscription helping generate data. Training took a little over ten hours, with estimated hardware rental below US$100. The first version took around five days and the second two or three weeks. Labor, owning hardware, failed experiments and continuing service are not fully counted. The title cannot establish another team’s total budget.
1. Begin with the objective; a benchmark is one check
Episode reference: 04:52–11:24. Lu reports one agent benchmark rising from about 15 to 55–60 while acknowledging declines elsewhere. Agents are model systems using tools to execute tasks. The benchmark, scale and conditions are not independently verified, so these numbers cannot become a several-fold gain in general intelligence.
The proposed sequence is defining a task, choosing a base, constructing data, selecting a method, training, examining errors and revising data. SFT, supervised fine-tuning, trains behavior from supplied examples. It suits this personal budget without establishing superiority over reinforcement learning for every objective.
Base models also require language, architecture and deployment compatibility. Poor Chinese in one base is a version- and task-specific experience rather than a permanent judgment about a family. AI coding can modify scripts; execution does not establish clean evaluation, absence of leakage or production readiness.
Editorial interpretation: Define the user outcome before selecting technology. If prompts, tools or retrieval already suffice, training may not be the first bottleneck. Benchmark gains need checking against real examples excluded from training.
2. Data takes time because a failure is more specific than a slogan
Episode reference: 09:17–15:34. Lu estimates spending about 95% of the time on data, including roughly twelve days for the second version. Around sixty to seventy percent came from real traces, supplemented by targeted synthesis, with a few thousand examples after filtering. This is a project recipe rather than a universal ratio.
The specific failure was declaring completion before finishing a task. Simply requesting data to cure overconfidence can produce humble language without correcting tool-state and stopping judgments. Human checks must identify what is wrong and refine generation constraints.
Personally reviewing around five hundred of five thousand examples illustrates work rather than statistical sufficiency. Personal usage, public material and community contributions also do not automatically carry identical rights. Model outputs and user logs have specific usage scopes.
Editorial interpretation: Sample by provenance, task, risk and failure type. Connect tool receipts, final artifacts and completion statements rather than merely teaching “perhaps I failed.” Synthetic data cannot replace actual results, while real data cannot bypass privacy and permission.
3. Correct teacher traces may exceed what a student can learn
Episode reference: 15:34–19:56. Capacity gap describes a strong teacher using implicit knowledge, long reasoning or complex tools that a small student may imitate textually without gaining stable ability. Apparently good examples can still fail to improve training.
The teacher-parameter speculation is not reliable architecture evidence and is not reconstructed. Lack of improvement might also reflect formatting, learning settings, mixture or evaluation errors, rather than student size alone.
Lu judges distillation by its intended task. Training benchmark variants while intending to improve safety misses the objective even if scores rise. Target improvement likewise cannot excuse regression elsewhere. Confused transcription of on-policy terminology is unsuitable for a precise taxonomy.
Editorial interpretation: Preserve data, base, settings and evaluation versions, changing interpretable factors sequentially. Test target tasks alongside language, general capability and safety regressions. A stronger teacher is not the full answer; what the student actually learns is the acceptance question.
4. Training and serving need different accounts
Episode reference: 20:01–26:36. Asked about continued usage, Lu immediately distinguishes serving: deployment frameworks, architecture, concurrency and environments introduce more problems. An H200 serving a dozen employees is a rough illustration without rental or purchase terms, billing period, context or throughput. It cannot guide procurement.
Novel-writing model experiments emphasize tooling maturity. Lu found Qwen-related training and deployment easier than some less common architectures at the time. This may affect a choice without becoming a current ranking. Check working examples instead of trusting AI suggestions alone.
An integrated training service could reduce repeated scripting, yet one-click training supplies a candidate model. Platforms cannot replace data rights, task definitions or business acceptance.
Editorial interpretation: Include labor, data, iterations, deployment, peaks, idle capacity, monitoring and upgrades. Self-hosting may cost more at low or variable demand. Stable narrow tasks and sensitive data can produce different economics, but actual workloads must test them.
5. An application becoming a small lab needs learnable feedback
Episode reference: 23:44–36:50. The speakers imagine applications with users, scenarios and operating businesses doing targeted post-training, frontier labs retaining expensive pretraining and research, and smaller teams buying services. This is a future division-of-labor hypothesis, not a requirement for every application to hire three or four trainers immediately.
User records are not automatically supervision. Conversation, failed clicks and ambiguous preferences still need transformation into tasks, outcomes and verifiable feedback. More records are not valuable without filtering.
They worry about changes in open-model supply, while Qu expects competition to sustain openness. Concern does not establish that open releases are about to stop, nor require rushed training by every company.
Editorial interpretation: Applications may accumulate advantages in discovering real failures, obtaining legitimate data, validating improvements and returning them to users. Application and foundation-model value can grow together. A small experiment does not require first deciding which side eventually absorbs the other.
6. Local AI offers several values with different costs
Episode reference: 36:59–46:24. Local models may support sensitive data or expensive creative workloads while reducing transmission to model providers. Lu admits limited personal local usage and remaining errors and capability gaps. The producer asks about frontier capability; Qu asks whether ordinary users will migrate for privacy, adding convenience and application defaults to the discussion.
Local does not mean data never leave. Telemetry, plugins, logs, synchronization and tools may transmit information. Previously purchased devices also retain operating, maintenance and electricity costs. Future phone models, Apple memory and anonymous leakage anecdotes are unverified rather than promises or established accusations.
Editorial interpretation: Test privacy, offline availability, quality, latency, expense and convenience separately. Sensitive businesses may have stronger deployment reasons while mass users prioritize ease. Both can coexist without one audience deciding for everyone.
Record complete costs and acceptance for one fine-tuning experiment
For an editorial exercise, choose a stable narrow task and fix independent evaluation and regression sets. Compare an existing API, prompt-and-tool changes, retrieval and fine-tuning. Preserve failure categories, completion evidence, rights and versions, then count data work, iterations and serving costs.
If the target improves but review grows, previous abilities regress or total cost rises, investigate rather than substituting popularity for results. Discuss deployment after offline validation. This is an editorial method, not a training tutorial replicated for this article.
Listening guide and related reading
- 03:23: The cost boundary of the personal project.
- 06:08: From objective to training.
- 11:33: Data and premature completion claims.
- 15:34: Capacity gaps.
- 20:01: Training versus serving costs.
- 27:01: Applications as labs, a hypothesis.
- 40:17: Frontier capability, locality and convenience.
Related themes: small models, iterative data and complete costs. Compare Meng on data quality and Zhu on actual operating costs, examining how training results become service outcomes.
Source boundary: The complete supplied transcript, earlier note and official introduction were read without sentence-by-sentence audio, model-card inspection, billing checks or training. Popularity and performance are separated; costs and scores remain Lu’s reports. Unclear bases, teachers, serving configurations and future hardware are not reconstructed into reproduction guarantees or purchase recommendations.