{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"从网页文字到训练对应：数据扩大了什么","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.md","pages":[{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","unit":"规模化图文监督带来了哪些可验证的进展？","unitId":"clip-data-and-transfer","chapter":"III","number":1}],"aliases":[{"oldId":"b-6fb26b3ebcc56dbea8be3395","targetId":"b-6fb26b3ebcc56dbea8be3395","page":"reading-opening","operation":"move","beforeSHA256":"364b26b33d8efc603eeafa55f5a4971dd84b28b1b54c77c7043133e1a4487685","afterSHA256":"364b26b33d8efc603eeafa55f5a4971dd84b28b1b54c77c7043133e1a4487685","reason":"Same source body in authored semantic page"},{"oldId":"b-c60c4bb58bcd390a86c5f792","targetId":"b-c60c4bb58bcd390a86c5f792","page":"reading-opening","operation":"move","beforeSHA256":"460d22b03b1e3a2d67fb3cc29b83b944a418620d3baef8805915d48014e91963","afterSHA256":"460d22b03b1e3a2d67fb3cc29b83b944a418620d3baef8805915d48014e91963","reason":"Same source body in authored semantic page"},{"oldId":"b-83008c61b157084e0e5de7a8","targetId":"b-83008c61b157084e0e5de7a8","page":"reading-opening","operation":"move","beforeSHA256":"ec4cdf9970d4040758e5e459fda33f9756daae8d7d4deb990af587e9df8df8a4","afterSHA256":"ec4cdf9970d4040758e5e459fda33f9756daae8d7d4deb990af587e9df8df8a4","reason":"Same source body in authored semantic page"},{"oldId":"b-d7fea6929e53d0eded3a1083","targetId":"b-d7fea6929e53d0eded3a1083","page":"reading-opening","operation":"move","beforeSHA256":"9cf21ed0941cab251857befbc9abc4b494b382f976b9ef11caaf27b3aeb0f77b","afterSHA256":"9cf21ed0941cab251857befbc9abc4b494b382f976b9ef11caaf27b3aeb0f77b","reason":"Same source body in authored semantic page"},{"oldId":"b-bf9933f1d8780a6926dd67bf","targetId":"b-bf9933f1d8780a6926dd67bf","page":"reading-opening","operation":"move","beforeSHA256":"434298b61a004a1f6cd3dcfa02b72a25c70c8991c129994ce20ece43594dfb5e","afterSHA256":"434298b61a004a1f6cd3dcfa02b72a25c70c8991c129994ce20ece43594dfb5e","reason":"Same source body in authored semantic page"},{"oldId":"b-61a189c64a4bc9187f89d321","targetId":"b-61a189c64a4bc9187f89d321","page":"reading-opening","operation":"move","beforeSHA256":"0a05e9cb8f047833928af47ddc13e86af6874f3533e85e00df2028121fc3b264","afterSHA256":"0a05e9cb8f047833928af47ddc13e86af6874f3533e85e00df2028121fc3b264","reason":"Same source body in authored semantic page"},{"oldId":"b-025fa91d652ba5154a561b80","targetId":"b-025fa91d652ba5154a561b80","page":"reading-opening","operation":"rewrite","beforeSHA256":"b73f7703988fa6f00378d358341c4fc1126dfba7005f5230498027fe8f048bee","afterSHA256":"59247b5dc71cbbda09c9536bce686a551fcd160debd38b35159fe9661f65855d","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-3979c411d1dececa93bdd145","targetId":"b-3979c411d1dececa93bdd145","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-6fb26b3ebcc56dbea8be3395","type":"h1","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-6fb26b3ebcc56dbea8be3395","content":"# 从网页文字到训练对应：数据扩大了什么","sha256":"364b26b33d8efc603eeafa55f5a4971dd84b28b1b54c77c7043133e1a4487685","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-c60c4bb58bcd390a86c5f792","prerequisites":[{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-c60c4bb58bcd390a86c5f792","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-c60c4bb58bcd390a86c5f792","content":"大规模图文监督的机会来自配对供给，信号质量又取决于描述、选择和筛选。原始 CLIP、LAION 和 OpenCLIP 分属模型研究、数据资源与实现/模型系列，不能互当别名。正面的迁移证据与数据边界一起读，才能理解这种路线为何有效、又在哪些地方需要下一步工作。","sha256":"460d22b03b1e3a2d67fb3cc29b83b944a418620d3baef8805915d48014e91963","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6fb26b3ebcc56dbea8be3395","nextBlock":"b-83008c61b157084e0e5de7a8","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-83008c61b157084e0e5de7a8","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-83008c61b157084e0e5de7a8","content":"网页配图旁的文字可能准确描述对象，也可能只讨论文章主题。图片中有两个杯子，文字却只提其中一个；图文仍相关，却没有提供完整场景标注。这种监督更便于规模化收集，但目标覆盖范围由内容生产与筛选决定。","sha256":"ec4cdf9970d4040758e5e459fda33f9756daae8d7d4deb990af587e9df8df8a4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c60c4bb58bcd390a86c5f792","nextBlock":"b-d7fea6929e53d0eded3a1083","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-d7fea6929e53d0eded3a1083","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-d7fea6929e53d0eded3a1083","content":"原始 CLIP 工作收集了 4 亿图文对，并在 30 多个既有视觉数据集上研究迁移。LAION-400M 是后来的开放图文数据资源，包含经 CLIP 过滤的图文对；它不是原始 CLIP 的训练集。OpenCLIP 又是另一层的实现和模型系列，具体权重使用什么数据要逐项核对。","sha256":"9cf21ed0941cab251857befbc9abc4b494b382f976b9ef11caaf27b3aeb0f77b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-83008c61b157084e0e5de7a8","nextBlock":"b-bf9933f1d8780a6926dd67bf","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-bf9933f1d8780a6926dd67bf","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-bf9933f1d8780a6926dd67bf","content":"这里存在一个积极反馈：已有模型可以帮助整理更多配对，开放的数据和实现使更多研究者能训练与比较新模型。但筛选器偏好的对应也可能进入数据分布。数量增加提供更多学习机会；对象、语言、关系、质量和去重方式决定这些机会覆盖什么。","sha256":"434298b61a004a1f6cd3dcfa02b72a25c70c8991c129994ce20ece43594dfb5e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d7fea6929e53d0eded3a1083","nextBlock":"b-61a189c64a4bc9187f89d321","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-61a189c64a4bc9187f89d321","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-61a189c64a4bc9187f89d321","content":"回到第一部分的标签供给，人的工作并没有消失，只是有的从逐图填写固定类别，转向写内容、选择资料、定义任务和评价。研究报告需要同时交代模型、数据与接口；单说“用了 CLIP/LAION”还不足以重建学习条件。","sha256":"0a05e9cb8f047833928af47ddc13e86af6874f3533e85e00df2028121fc3b264","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bf9933f1d8780a6926dd67bf","nextBlock":"b-025fa91d652ba5154a561b80","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]},{"id":"b-025fa91d652ba5154a561b80","type":"figure","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-025fa91d652ba5154a561b80","content":"图内标签与数值：“两个人站在室内。”；收集与筛选；保留哪些配对？；图文训练；配对关系提供目标；筛选模型影响材料范围；CLIP；模型与训练方法；原始数据由原工作收集；LAION；公开图文数据资源；不是原始 CLIP 训练集；OpenCLIP；开源实现与模型；不同权重有不同数据配方；人工内容、数据处理与筛选都参与信号形成；无额外逐图类别标签并不等于无人类知识。\n\n同一名称不能同时代替模型、数据资源和实现；学习信号需要追溯来源。 查看对应图解状态\n\n- “两个人站在室内。”\n- 收集与筛选\n- 保留哪些配对？\n- 图文训练\n- 配对关系提供目标\n- 筛选模型影响材料范围\n- CLIP\n- 模型与训练方法\n- 原始数据由原工作收集\n- LAION\n- 公开图文数据资源\n- 不是原始 CLIP 训练集\n- OpenCLIP\n- 开源实现与模型\n- 不同权重有不同数据配方\n- 人工内容、数据处理与筛选都参与信号形成；无额外逐图类别标签并不等于无人类知识。同一名称不能同时代替模型、数据资源和实现；学习信号需要追溯来源。 [查看对应图解状态](/#/l3-r2-p60/0)[查看完整图解](/#/l3-r2-p60/0)","sha256":"59247b5dc71cbbda09c9536bce686a551fcd160debd38b35159fe9661f65855d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一名称不能同时代替模型、数据资源和实现；学习信号需要追溯来源。 查看对应图解状态","labels":["“两个人站在室内。”","收集与筛选","保留哪些配对？","图文训练","配对关系提供目标","筛选模型影响材料范围","CLIP","模型与训练方法","原始数据由原工作收集","LAION","公开图文数据资源","不是原始 CLIP 训练集","OpenCLIP","开源实现与模型","不同权重有不同数据配方","人工内容、数据处理与筛选都参与信号形成；无额外逐图类别标签并不等于无人类知识。"],"description":"同一名称不能同时代替模型、数据资源和实现；学习信号需要追溯来源。 查看对应图解状态“两个人站在室内。”收集与筛选保留哪些配对？图文训练配对关系提供目标筛选模型影响材料范围CLIP模型与训练方法原始数据由原工作收集LAION公开图文数据资源不是原始 CLIP 训练集OpenCLIP开源实现与模型不同权重有不同数据配方人工内容、数据处理与筛选都参与信号形成；无额外逐图类别标签并不等于无人类知识。","evidenceIdentity":"同一名称不能同时代替模型、数据资源和实现；学习信号需要追溯来源。 查看对应图解状态","images":[{"src":"/assets/display/512610a019c560ff0384e86a.webp","alt":""}]},"previousBlock":"b-61a189c64a4bc9187f89d321","nextBlock":"b-3979c411d1dececa93bdd145","prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3979c411d1dececa93bdd145","type":"h2","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/blocks/clip-supervision-supply.html#b-3979c411d1dececa93bdd145","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-data-and-transfer","title":"规模化图文监督带来了哪些可验证的进展？"},"page":{"id":"reading-opening","title":"从网页文字到训练对应：数据扩大了什么","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"从网页文字到训练对应：数据扩大了什么","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-025fa91d652ba5154a561b80","nextBlock":null,"prerequisites":[{"id":"b-6fb26b3ebcc56dbea8be3395","url":"#b-6fb26b3ebcc56dbea8be3395","role":"section-title"},{"id":"b-c60c4bb58bcd390a86c5f792","url":"#b-c60c4bb58bcd390a86c5f792","role":"problem-context"}],"relatedBlocks":[{"id":"b-025fa91d652ba5154a561b80","url":"#b-025fa91d652ba5154a561b80","relation":"figure-and-explanation"}]}]}