{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"图文数据不是一堆没有来历的图片","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.md","pages":[{"id":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","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},{"id":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","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":2}],"aliases":[{"oldId":"explanation","targetId":"clip-supervision-provenance","operation":"split-explanation-range"},{"oldId":"b-5e78dfeec9ec93bbd3110749","targetId":"b-5e78dfeec9ec93bbd3110749","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-ca6e2188e4545856b1160796","targetId":"b-ca6e2188e4545856b1160796","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"ebecb10875bdc61bfb921d89f57842b123e6b16dd1477ca74b66646c169bd0ee","afterSHA256":"ebecb10875bdc61bfb921d89f57842b123e6b16dd1477ca74b66646c169bd0ee","reason":"Same source body in authored semantic page"},{"oldId":"b-66da3f52977d09daf5571382","targetId":"b-66da3f52977d09daf5571382","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"460d22b03b1e3a2d67fb3cc29b83b944a418620d3baef8805915d48014e91963","afterSHA256":"460d22b03b1e3a2d67fb3cc29b83b944a418620d3baef8805915d48014e91963","reason":"Same source body in authored semantic page"},{"oldId":"b-df44c3091b487e6396ebb49f","targetId":"b-df44c3091b487e6396ebb49f","page":"clip-supervision-provenance","operation":"rewrite","beforeSHA256":"7467859cb5327496f9647c8c90b5a7095c6390e083cc350f113de4b0474c23d0","afterSHA256":"ef9cb0a3c1069c1cefbbad49cb06c4cb760e25e66edf262b1ec4d560f555d9f5","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-7176ac036d07d81bf2342f2a","targetId":"b-7176ac036d07d81bf2342f2a","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-1a6c595f809981bc32cf5add","targetId":"b-1a6c595f809981bc32cf5add","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"eab8bc54377ec85bcfe39888a37acbc6e90238d16803d354cb62777fd78adb0c","afterSHA256":"eab8bc54377ec85bcfe39888a37acbc6e90238d16803d354cb62777fd78adb0c","reason":"Same source body in authored semantic page"},{"oldId":"b-6dd9d05e9c90b705593134f4","targetId":"b-6dd9d05e9c90b705593134f4","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"27e1887db329033a1e332272fc447093911919abde345b1d59d39b8c342f004d","afterSHA256":"27e1887db329033a1e332272fc447093911919abde345b1d59d39b8c342f004d","reason":"Same source body in authored semantic page"},{"oldId":"b-ed71848c60d9095e41c3c935","targetId":"b-ed71848c60d9095e41c3c935","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"ec4cdf9970d4040758e5e459fda33f9756daae8d7d4deb990af587e9df8df8a4","afterSHA256":"ec4cdf9970d4040758e5e459fda33f9756daae8d7d4deb990af587e9df8df8a4","reason":"Same source body in authored semantic page"},{"oldId":"b-071ad7f17095a7b585cd6d0a","targetId":"b-071ad7f17095a7b585cd6d0a","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"9cf21ed0941cab251857befbc9abc4b494b382f976b9ef11caaf27b3aeb0f77b","afterSHA256":"9cf21ed0941cab251857befbc9abc4b494b382f976b9ef11caaf27b3aeb0f77b","reason":"Same source body in authored semantic page"},{"oldId":"b-08fa8d6400e366020b0635e0","targetId":"b-08fa8d6400e366020b0635e0","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"434298b61a004a1f6cd3dcfa02b72a25c70c8991c129994ce20ece43594dfb5e","afterSHA256":"434298b61a004a1f6cd3dcfa02b72a25c70c8991c129994ce20ece43594dfb5e","reason":"Same source body in authored semantic page"},{"oldId":"b-9181fc5e1a1a1f4986d95dea","targetId":"b-9181fc5e1a1a1f4986d95dea","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"0a05e9cb8f047833928af47ddc13e86af6874f3533e85e00df2028121fc3b264","afterSHA256":"0a05e9cb8f047833928af47ddc13e86af6874f3533e85e00df2028121fc3b264","reason":"Same source body in authored semantic page"},{"oldId":"b-9cd117ff4c129a8c00a622c1","targetId":"b-9cd117ff4c129a8c00a622c1","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"68745b5a79d05357c8edeeff919a146e1a56cd1890bfc603486d2a7e3e2326e6","afterSHA256":"68745b5a79d05357c8edeeff919a146e1a56cd1890bfc603486d2a7e3e2326e6","reason":"Same source body in authored semantic page"},{"oldId":"b-e8d4eef8ee8a364502782ae4","targetId":"b-e8d4eef8ee8a364502782ae4","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","afterSHA256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","reason":"Same source body in authored semantic page"},{"oldId":"b-1fd16cf1924d792a95ae49c1","targetId":"b-1fd16cf1924d792a95ae49c1","page":"clip-supervision-provenance","operation":"move","beforeSHA256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","afterSHA256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","reason":"Same source body in authored semantic page"},{"oldId":"b-1366bfae522c88e207520b4a","targetId":"b-1366bfae522c88e207520b4a","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"32fba3360fa674f66fececa5b66b13a3802216b91567c9dd5c1681cce7f2c586","afterSHA256":"32fba3360fa674f66fececa5b66b13a3802216b91567c9dd5c1681cce7f2c586","reason":"Same source body in authored semantic page"},{"oldId":"b-1919519921a42604b83dca8f","targetId":"b-1919519921a42604b83dca8f","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"6d7f0dd74b4667c6a08b0646dce855f483d825be4b813261f23dc52e0afb3b1f","afterSHA256":"6d7f0dd74b4667c6a08b0646dce855f483d825be4b813261f23dc52e0afb3b1f","reason":"Same source body in authored semantic page"},{"oldId":"b-a9830f247b5f0262e7214629","targetId":"b-a9830f247b5f0262e7214629","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"1212618a46874991135f8b45069f106d46e0cb3f6ac860805804ed731db72c91","afterSHA256":"1212618a46874991135f8b45069f106d46e0cb3f6ac860805804ed731db72c91","reason":"Same source body in authored semantic page"},{"oldId":"b-6b8b88919924c34e4b1485d0","targetId":"b-6b8b88919924c34e4b1485d0","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"1ac2ccaaed2600cd171432e84791b757c5c816d12897751f508cab63ff1af1dc","afterSHA256":"1ac2ccaaed2600cd171432e84791b757c5c816d12897751f508cab63ff1af1dc","reason":"Same source body in authored semantic page"},{"oldId":"b-7a6d0d4f2ae788cd0d334767","targetId":"b-7a6d0d4f2ae788cd0d334767","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-12518081eea47c8ac2f88d94","targetId":"b-12518081eea47c8ac2f88d94","page":"clip-transfer-evidence","operation":"move","beforeSHA256":"69e0a6a8a30bd23bedb216d03f4786b7b1d601ac3b0b6576fc326a5d3e873aae","afterSHA256":"69e0a6a8a30bd23bedb216d03f4786b7b1d601ac3b0b6576fc326a5d3e873aae","reason":"Same source body in authored semantic page"},{"oldId":"clip-supervision-provenance-title","targetId":"b-ca7378c2cfe36649753fb7fc","operation":"retain-source-identity"},{"oldId":"clip-transfer-evidence-title","targetId":"b-05a9d56a4a3543bd5ea2630d","operation":"retain-source-identity"},{"oldId":"clip-transfer-evidence-bridge","targetId":"b-d376b1fa3aacd507acca4877","operation":"retain-source-identity"}],"blocks":[{"id":"b-ca7378c2cfe36649753fb7fc","type":"h2","scope":"clip-supervision-provenance","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-ca7378c2cfe36649753fb7fc","content":"## 追溯配对供给与筛选覆盖","sha256":"8760fca919a9e3b5d89d603c6b41a776f117f6956d022e6dd95feeaf44f9ac11","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5e78dfeec9ec93bbd3110749","prerequisites":[{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-5e78dfeec9ec93bbd3110749","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-5e78dfeec9ec93bbd3110749","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ca7378c2cfe36649753fb7fc","nextBlock":"b-ca6e2188e4545856b1160796","prerequisites":[{"id":"b-ca7378c2cfe36649753fb7fc","url":"#b-ca7378c2cfe36649753fb7fc","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-ca6e2188e4545856b1160796","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-ca6e2188e4545856b1160796","content":"# P60 · 图文数据不是一堆没有来历的图片","sha256":"ebecb10875bdc61bfb921d89f57842b123e6b16dd1477ca74b66646c169bd0ee","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5e78dfeec9ec93bbd3110749","nextBlock":"b-66da3f52977d09daf5571382","prerequisites":[{"id":"b-5e78dfeec9ec93bbd3110749","url":"#b-5e78dfeec9ec93bbd3110749","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-66da3f52977d09daf5571382","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-66da3f52977d09daf5571382","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ca6e2188e4545856b1160796","nextBlock":"b-df44c3091b487e6396ebb49f","prerequisites":[{"id":"b-ca6e2188e4545856b1160796","url":"#b-ca6e2188e4545856b1160796","role":"section-title"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-df44c3091b487e6396ebb49f","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-df44c3091b487e6396ebb49f","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原图与描述经过收集筛选进入训练；下方分别说明CLIP模型方法、LAION公开数据和OpenCLIP开源实现。\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。原图与描述经过收集筛选进入训练；下方分别说明CLIP模型方法、LAION公开数据和OpenCLIP开源实现。 [查看对应图解状态](/#/l3-r2-p60/0)[查看完整图解](/course/figures/l3-r2-p60-0.html)","sha256":"ef9cb0a3c1069c1cefbbad49cb06c4cb760e25e66edf262b1ec4d560f555d9f5","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"原图与描述经过收集筛选进入训练；下方分别说明CLIP模型方法、LAION公开数据和OpenCLIP开源实现。 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":"原图与描述经过收集筛选进入训练；下方分别说明CLIP模型方法、LAION公开数据和OpenCLIP开源实现。“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":"原图与描述经过收集筛选进入训练；下方分别说明CLIP模型方法、LAION公开数据和OpenCLIP开源实现。 查看对应图解状态","images":[{"src":"/assets/display/512610a019c560ff0384e86a.webp","alt":""}]},"previousBlock":"b-66da3f52977d09daf5571382","nextBlock":"b-7176ac036d07d81bf2342f2a","prerequisites":[{"id":"b-ca6e2188e4545856b1160796","url":"#b-ca6e2188e4545856b1160796","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7176ac036d07d81bf2342f2a","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-7176ac036d07d81bf2342f2a","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-df44c3091b487e6396ebb49f","nextBlock":"b-1a6c595f809981bc32cf5add","prerequisites":[{"id":"b-ca6e2188e4545856b1160796","url":"#b-ca6e2188e4545856b1160796","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-1a6c595f809981bc32cf5add","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-1a6c595f809981bc32cf5add","content":"语言监督的规模，依赖图文配对数据怎样获得和整理。图片旁边的文字可能是描述，也可能只说了背景或网站主题；配对质量不是天然一致的。\nLAION 等公开资源使更多研究者能够研究大规模图文学习，但它不是原始 CLIP 的训练集。模型、训练数据和开源实现要分开说，不能把几个名字当同一个东西。\n这也回到第一部分的监督供给问题。人不仅可能直接标答案，还可能通过内容生产、筛选与评价影响模型学到什么。数据更多扩大了机会，数据如何组织则决定这些机会里包含什么信号。理解这种条件，才不会把整个变化只归结为参数数量。","sha256":"eab8bc54377ec85bcfe39888a37acbc6e90238d16803d354cb62777fd78adb0c","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7176ac036d07d81bf2342f2a","nextBlock":"b-6dd9d05e9c90b705593134f4","prerequisites":[{"id":"b-7176ac036d07d81bf2342f2a","url":"#b-7176ac036d07d81bf2342f2a","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-6dd9d05e9c90b705593134f4","type":"h3","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-6dd9d05e9c90b705593134f4","content":"### 从网页文字到训练对应：数据扩大了什么","sha256":"27e1887db329033a1e332272fc447093911919abde345b1d59d39b8c342f004d","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":"clip-supervision-supply","formulas":[],"previousBlock":"b-1a6c595f809981bc32cf5add","nextBlock":"b-ed71848c60d9095e41c3c935","prerequisites":[{"id":"b-7176ac036d07d81bf2342f2a","url":"#b-7176ac036d07d81bf2342f2a","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"},{"id":"b-ed71848c60d9095e41c3c935","url":"#b-ed71848c60d9095e41c3c935","relation":"same-reasoning"},{"id":"b-071ad7f17095a7b585cd6d0a","url":"#b-071ad7f17095a7b585cd6d0a","relation":"same-reasoning"},{"id":"b-08fa8d6400e366020b0635e0","url":"#b-08fa8d6400e366020b0635e0","relation":"same-reasoning"},{"id":"b-9181fc5e1a1a1f4986d95dea","url":"#b-9181fc5e1a1a1f4986d95dea","relation":"same-reasoning"}]},{"id":"b-ed71848c60d9095e41c3c935","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-ed71848c60d9095e41c3c935","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":"clip-supervision-supply","formulas":[],"previousBlock":"b-6dd9d05e9c90b705593134f4","nextBlock":"b-071ad7f17095a7b585cd6d0a","prerequisites":[{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"},{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","relation":"same-reasoning"},{"id":"b-071ad7f17095a7b585cd6d0a","url":"#b-071ad7f17095a7b585cd6d0a","relation":"same-reasoning"},{"id":"b-08fa8d6400e366020b0635e0","url":"#b-08fa8d6400e366020b0635e0","relation":"same-reasoning"},{"id":"b-9181fc5e1a1a1f4986d95dea","url":"#b-9181fc5e1a1a1f4986d95dea","relation":"same-reasoning"}]},{"id":"b-071ad7f17095a7b585cd6d0a","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-071ad7f17095a7b585cd6d0a","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":"clip-supervision-supply","formulas":[],"previousBlock":"b-ed71848c60d9095e41c3c935","nextBlock":"b-08fa8d6400e366020b0635e0","prerequisites":[{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"},{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","relation":"same-reasoning"},{"id":"b-ed71848c60d9095e41c3c935","url":"#b-ed71848c60d9095e41c3c935","relation":"same-reasoning"},{"id":"b-08fa8d6400e366020b0635e0","url":"#b-08fa8d6400e366020b0635e0","relation":"same-reasoning"},{"id":"b-9181fc5e1a1a1f4986d95dea","url":"#b-9181fc5e1a1a1f4986d95dea","relation":"same-reasoning"}]},{"id":"b-08fa8d6400e366020b0635e0","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-08fa8d6400e366020b0635e0","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":"clip-supervision-supply","formulas":[],"previousBlock":"b-071ad7f17095a7b585cd6d0a","nextBlock":"b-9181fc5e1a1a1f4986d95dea","prerequisites":[{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"},{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","relation":"same-reasoning"},{"id":"b-ed71848c60d9095e41c3c935","url":"#b-ed71848c60d9095e41c3c935","relation":"same-reasoning"},{"id":"b-071ad7f17095a7b585cd6d0a","url":"#b-071ad7f17095a7b585cd6d0a","relation":"same-reasoning"},{"id":"b-9181fc5e1a1a1f4986d95dea","url":"#b-9181fc5e1a1a1f4986d95dea","relation":"same-reasoning"}]},{"id":"b-9181fc5e1a1a1f4986d95dea","type":"p","scope":"clip-supervision-supply","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-9181fc5e1a1a1f4986d95dea","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"从网页文字到训练对应：数据扩大了什么","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":"clip-supervision-supply","formulas":[],"previousBlock":"b-08fa8d6400e366020b0635e0","nextBlock":"b-98fc4f432304cd4c69bf7da8","prerequisites":[{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"},{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","relation":"same-reasoning"},{"id":"b-ed71848c60d9095e41c3c935","url":"#b-ed71848c60d9095e41c3c935","relation":"same-reasoning"},{"id":"b-071ad7f17095a7b585cd6d0a","url":"#b-071ad7f17095a7b585cd6d0a","relation":"same-reasoning"},{"id":"b-08fa8d6400e366020b0635e0","url":"#b-08fa8d6400e366020b0635e0","relation":"same-reasoning"}]},{"id":"b-98fc4f432304cd4c69bf7da8","type":"h4","scope":"clip-supervision-provenance","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-98fc4f432304cd4c69bf7da8","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9181fc5e1a1a1f4986d95dea","nextBlock":"b-9cd117ff4c129a8c00a622c1","prerequisites":[{"id":"b-6dd9d05e9c90b705593134f4","url":"#b-6dd9d05e9c90b705593134f4","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-9cd117ff4c129a8c00a622c1","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-9cd117ff4c129a8c00a622c1","content":"原始 CLIP 的训练数据由原工作收集；LAION 是后来的公开图文数据资源，OpenCLIP 是独立的开源实现与模型系列。具体权重使用的训练数据要分别核对。","sha256":"68745b5a79d05357c8edeeff919a146e1a56cd1890bfc603486d2a7e3e2326e6","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98fc4f432304cd4c69bf7da8","nextBlock":"b-546aa3dd892f6fbe180c9af0","prerequisites":[{"id":"b-98fc4f432304cd4c69bf7da8","url":"#b-98fc4f432304cd4c69bf7da8","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-546aa3dd892f6fbe180c9af0","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-546aa3dd892f6fbe180c9af0","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9cd117ff4c129a8c00a622c1","nextBlock":"b-e8d4eef8ee8a364502782ae4","prerequisites":[{"id":"b-98fc4f432304cd4c69bf7da8","url":"#b-98fc4f432304cd4c69bf7da8","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-e8d4eef8ee8a364502782ae4","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-e8d4eef8ee8a364502782ae4","content":"- [网络图文扩大了信号来源，也带来选择与噪声](/supplement.html?return=l3-r2-p60/0#/l3-clip-data/0) 本页说明","sha256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-546aa3dd892f6fbe180c9af0","nextBlock":"b-1fd16cf1924d792a95ae49c1","prerequisites":[{"id":"b-546aa3dd892f6fbe180c9af0","url":"#b-546aa3dd892f6fbe180c9af0","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-1fd16cf1924d792a95ae49c1","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-1fd16cf1924d792a95ae49c1","content":"- [网络图文扩大了信号来源，也带来选择与噪声](/supplement.html?return=l3-r2-p60/0#/l3-clip-data/0) 本页说明","sha256":"fc4388afc9540d40a4237dcc9d7454676e613008528c50f53547fb90e3cb5ff6","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":"clip-supervision-provenance","title":"追溯配对供给与筛选覆盖","number":1},"sectionTitle":"追溯配对供给与筛选覆盖","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e8d4eef8ee8a364502782ae4","nextBlock":null,"prerequisites":[{"id":"b-546aa3dd892f6fbe180c9af0","url":"#b-546aa3dd892f6fbe180c9af0","role":"section-title"},{"id":"b-66da3f52977d09daf5571382","url":"#b-66da3f52977d09daf5571382","role":"problem-context"}],"relatedBlocks":[{"id":"b-df44c3091b487e6396ebb49f","url":"#b-df44c3091b487e6396ebb49f","relation":"figure-and-explanation"}]},{"id":"b-05a9d56a4a3543bd5ea2630d","type":"h2","scope":"clip-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-05a9d56a4a3543bd5ea2630d","content":"## 先确认语言指定任务的实质进展","sha256":"e4a5808ead3032876e5ec8b808f49796945f5e7a946069472d69c75bb186beec","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"先确认语言指定任务的实质进展","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-d376b1fa3aacd507acca4877","prerequisites":[{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d376b1fa3aacd507acca4877","type":"p","scope":"clip-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-d376b1fa3aacd507acca4877","content":"规模提供机会，但进展还要由明确协议下的迁移结果支持。","sha256":"8017403ed296a6584280fd34991b8e1dfb05089367d532635c085af4b1fcff29","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"先确认语言指定任务的实质进展","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-05a9d56a4a3543bd5ea2630d","nextBlock":"b-1366bfae522c88e207520b4a","prerequisites":[{"id":"b-05a9d56a4a3543bd5ea2630d","url":"#b-05a9d56a4a3543bd5ea2630d","role":"section-title"}],"relatedBlocks":[]},{"id":"b-1366bfae522c88e207520b4a","type":"h3","scope":"clip-positive-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-1366bfae522c88e207520b4a","content":"### 正面证据：自然语言监督确实拓宽了下游接口","sha256":"32fba3360fa674f66fececa5b66b13a3802216b91567c9dd5c1681cce7f2c586","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"正面证据：自然语言监督确实拓宽了下游接口","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":"clip-positive-transfer-evidence","formulas":[],"previousBlock":"b-d376b1fa3aacd507acca4877","nextBlock":"b-1919519921a42604b83dca8f","prerequisites":[{"id":"b-05a9d56a4a3543bd5ea2630d","url":"#b-05a9d56a4a3543bd5ea2630d","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[{"id":"b-1919519921a42604b83dca8f","url":"#b-1919519921a42604b83dca8f","relation":"same-reasoning"},{"id":"b-a9830f247b5f0262e7214629","url":"#b-a9830f247b5f0262e7214629","relation":"same-reasoning"},{"id":"b-6b8b88919924c34e4b1485d0","url":"#b-6b8b88919924c34e4b1485d0","relation":"same-reasoning"}]},{"id":"b-1919519921a42604b83dca8f","type":"p","scope":"clip-positive-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-1919519921a42604b83dca8f","content":"CLIP 的论文贡献不只是一种损失公式。原工作研究了 30 多个视觉数据集上的迁移，且报告其 zero-shot 模型在 ImageNet 上达到原始 ResNet-50 的分类准确率水平，而不使用该任务的 128 万张训练样本训练专用分类器。数字与比较属于原论文协议，不能从本课程猫图实验重新推出。","sha256":"6d7f0dd74b4667c6a08b0646dce855f483d825be4b813261f23dc52e0afb3b1f","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"正面证据：自然语言监督确实拓宽了下游接口","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":"clip-positive-transfer-evidence","formulas":[],"previousBlock":"b-1366bfae522c88e207520b4a","nextBlock":"b-a9830f247b5f0262e7214629","prerequisites":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","relation":"same-reasoning"},{"id":"b-a9830f247b5f0262e7214629","url":"#b-a9830f247b5f0262e7214629","relation":"same-reasoning"},{"id":"b-6b8b88919924c34e4b1485d0","url":"#b-6b8b88919924c34e4b1485d0","relation":"same-reasoning"}]},{"id":"b-a9830f247b5f0262e7214629","type":"p","scope":"clip-positive-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-a9830f247b5f0262e7214629","content":"这项结果支持一个明确判断：大规模图文匹配学到的表示，可以通过文字描述定义一批下游分类任务，减少逐任务重新标注和训练固定分类头的需求。语言成为任务指定接口，是对“分类表只能在训练时固定”的实质拓展。","sha256":"1212618a46874991135f8b45069f106d46e0cb3f6ac860805804ed731db72c91","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"正面证据：自然语言监督确实拓宽了下游接口","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":"clip-positive-transfer-evidence","formulas":[],"previousBlock":"b-1919519921a42604b83dca8f","nextBlock":"b-6b8b88919924c34e4b1485d0","prerequisites":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","relation":"same-reasoning"},{"id":"b-1919519921a42604b83dca8f","url":"#b-1919519921a42604b83dca8f","relation":"same-reasoning"},{"id":"b-6b8b88919924c34e4b1485d0","url":"#b-6b8b88919924c34e4b1485d0","relation":"same-reasoning"}]},{"id":"b-6b8b88919924c34e4b1485d0","type":"p","scope":"clip-positive-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-6b8b88919924c34e4b1485d0","content":"然后再问边界才有意义：候选描述写得好不好、预训练是否覆盖概念、目标数据分布是否变化，都会影响迁移。关系定位和自由生成还要求别的输出。不是因为 CLIP 什么也没有证明而要继续研究，而是它已经打开了可用接口，研究者因此可以追问更细的空间、关系和交互能力。","sha256":"1ac2ccaaed2600cd171432e84791b757c5c816d12897751f508cab63ff1af1dc","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"正面证据：自然语言监督确实拓宽了下游接口","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":"clip-positive-transfer-evidence","formulas":[],"previousBlock":"b-a9830f247b5f0262e7214629","nextBlock":"b-7a6d0d4f2ae788cd0d334767","prerequisites":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","relation":"same-reasoning"},{"id":"b-1919519921a42604b83dca8f","url":"#b-1919519921a42604b83dca8f","relation":"same-reasoning"},{"id":"b-a9830f247b5f0262e7214629","url":"#b-a9830f247b5f0262e7214629","relation":"same-reasoning"}]},{"id":"b-7a6d0d4f2ae788cd0d334767","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-7a6d0d4f2ae788cd0d334767","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"先确认语言指定任务的实质进展","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6b8b88919924c34e4b1485d0","nextBlock":"b-12518081eea47c8ac2f88d94","prerequisites":[{"id":"b-1366bfae522c88e207520b4a","url":"#b-1366bfae522c88e207520b4a","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-12518081eea47c8ac2f88d94","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p60.html#b-12518081eea47c8ac2f88d94","content":"先区分原始 CLIP、LAION 与 OpenCLIP 的身份，再比较配对收集、筛选和模型训练各自改变什么。结合迁移结果，说明文字监督扩大了哪些任务接口。","sha256":"69e0a6a8a30bd23bedb216d03f4786b7b1d601ac3b0b6576fc326a5d3e873aae","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":"clip-transfer-evidence","title":"先确认语言指定任务的实质进展","number":2},"sectionTitle":"先确认语言指定任务的实质进展","currentQuestion":"模型、训练数据与使用接口分别贡献了什么？","sourceSlides":["l3-r2-p56","l3-r2-p60"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7a6d0d4f2ae788cd0d334767","nextBlock":null,"prerequisites":[{"id":"b-7a6d0d4f2ae788cd0d334767","url":"#b-7a6d0d4f2ae788cd0d334767","role":"section-title"},{"id":"b-d376b1fa3aacd507acca4877","url":"#b-d376b1fa3aacd507acca4877","role":"problem-context"}],"relatedBlocks":[]}]}