llm-post-training skillA
llm-post-training is agent-read markdown (skill) from yuecao365/offercome: 大模型后训练深挖：RL 与奖励设计、推理模型训练、参数高效微调、数据合成与蒸馏、对齐税。.
Indexed from public GitHub and served as immutable, content-addressed versions. Install it pinned to an exact SHA-256 with the mdr CLI, and every file is verified against the hash recorded here before it reaches your agent. The deterministic audit below grades the latest version, and the same file always earns the same grade.
What the file says
## 面试官在意什么 这本是 ai-algorithm 里后训练那一段的深挖：SFT 之后模型怎么变强、怎么变可控。2026 年的面试重点是可验证奖励的 RL（GRPO、DAPO 与各家变体）、推理模型的训练细节、奖励怎么被 hack、合成数据怎么不把模型带偏。面试官会要求推导 DPO、写出 GRPO 目标里的裁剪与 KL 项、算 LoRA 的显存账，并追一次真实的训练事故：奖励 hack、长度膨胀、熵坍缩、评测污染。 ## 项目 / 实习怎么深挖 简历上出现下面这类经历时从哪里切、追什么。追到候选人能说出机制、数字的来源与一次真实的故障或取舍才算实；只有框架名与结论、说不出自己那一段的，记为危险信号。通用的追问方法见 project-deep-dive。 - 简历出现 RLHF / DPO / GRPO → 追偏好或奖励怎么来、有没有观察到 reward hacking、KL 与裁剪系数怎么调、和 SFT 基线涨在哪 - 简历出现推理模型 / 长思维链 → 追可验证奖励怎么设计与判分、长度膨胀怎么控、在不可验证任务上怎么迁移 - 简历出现 LoRA / QLoRA → 追 rank 与目标层、显存账、和全参对比、合并后有没有回归 - 简历出现数据合成 / 蒸馏 → 追教师模型、过滤规则、多样性与错误率怎么量、有没有验证过合成数据的收益 - 简历出现奖励模型 → 追训练数据、和人工一致性、过优化怎么发现 - 简历出现"评测涨 X 分" → 追污染检查、种子数、方差、对齐税（通用能力掉了多少） ## 常见失守与危险信号 - RL 与奖励设计：说不出 GRPO 与 PPO 的差别；不知道奖励可以被 hack；把"思维链更长"当能力更强 - 奖励模型与判分器：判分器只做字符串匹配；奖励模型没和人工对过；不知道过优化 - 推理模型训练：只会"用 R1 的方法"；说不出冷启动 SFT 的作用；长度与熵不一起看 - 参数高效微调：只知道 LoRA 的名字；算不出显存；合并权重后不做回归 - 数据合成与蒸馏：合成数据不过滤直接训；不知道分布坍缩；蒸馏只答"用大模型生成答案" - 数据配比与课程：配比拍脑袋；不知道通用数据抗遗忘；没做过消融 - 对齐税与回归：只看目标任务分数；不知道通用能力与安全性可能同时下降 - 在线与离线：分不清 on-policy 与 off-policy 对数据新鲜度的要求；DPO 数据来自别的模型却不觉得有问题 ## 常考主题清单 只列名字、阶梯与答实的标志，作"问到哪一层算实"的参考；问哪些、问几道由这份 JD 与这份简历定，不是配额。 ### RL for LLM：奖励设计与推理模型 …
Read the whole file at its exact version.
How to install
mdr add yuecao365/offercome/llm-post-training@git:20260920.a4843f5mdr add yuecao365/offercome/llm-post-training@sha256:ad34774b75739f79Pin to a label to follow the author's releases, or to a sha256 to freeze the exact bytes forever. Either way the resolved hash is written to mdr.lock, and mdr install reproduces it on any machine.
[](https://markdownregistry.com/a/art_g7r3bx7e7st4f4uk)
1 badge views in 30 days
Versions
Audit of the latest version
- pass: Frontmatter block present
- pass: Frontmatter declares a name
- pass: Frontmatter declares a description
- pass: Size between 200 bytes and 200 KB (7247 bytes)
- pass: No zero-width or bidi control characters
- pass: No instruction hidden inside an HTML comment
- pass: No link to an exfiltration or paste host
- pass: No credential-shaped string
- pass: No instruction to send local credentials anywhere
- pass: No text hidden with inline styles
- pass: No prompt-injection phrasing
- pass: No curl or wget piped into a shell
- pass: No recursive delete of root, home or parent
- pass: No instruction to read or print local credentials
- pass: No base64 blob over 200 characters
- pass: No link to a raw IP address
- pass: No script tag
Source
yuecao365/offercome · 23 stars · license MIT · pushed 2026-09-23 · branch main
API
GET https://markdownregistry.com/api/v1/artifacts/art_g7r3bx7e7st4f4uk GET https://markdownregistry.com/api/v1/resolve?ref=yuecao365/offercome/llm-post-training GET https://markdownregistry.com/api/v1/blob/ad34774b75739f79ca628899a889b473d350b3c3b9ae6bf80a14c344b62ea773
Your agent does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.