🧸 一家做儿童教育产品的公司把自己内部那张小学知识图谱整个开源了，项目叫 os-taxonomy，等于把「孩子该按什么顺序学什么」摊开给所有人用。

os-taxonomy 今年 7 月才建的仓库，现在 5881 star、986 个 fork，这个 fork 比例说明是真有人在拿它搭东西。

做教育类产品最头疼的不是界面，是知识结构。你想加一个自适应学习的功能，就得先回答「这孩子到底卡在哪一步」，而这需要一张完整的先修关系网——通常得请教研团队花几个月人工梳理，还未必梳理得全。

os-taxonomy 直接把这张网给了你：1590 个细粒度知识点、3221 条先修依赖，整理成有向无环图，每条边还标了是 hard 还是 soft 依赖并写明理由。覆盖科学、数学、英语、历史等 8 个学科，每个知识点都带稳定标识符、掌握证据标准、评估提示和大致适龄区间，并链接到 NGSS、Common Core、英国国家课程这些来源。另外还有 183 段专门写给家长看的领域总结。

许可也交代得很清楚：数据库部分 ODbL 1.0，Marble 自己写的文本部分 CC BY-SA 4.0，研究和商用都行，署名加上、衍生数据库保持开放就可以，基于它做出来的产品本身不用开源。这一版里没有嵌入向量，也没有任何孩子的个人数据。

教研的底稿被人无偿摊开，这事比多一个 App 有用。

GitHub：

withmarbleapp/os-taxonomy

Source: https://github.com/withmarbleapp/os-taxonomy

Marble Skill Taxonomy

An open, structured taxonomy of what children learn across the primary/elementary years — decomposed into fine-grained “micro-topics”, wired into a prerequisite graph, and aligned to national curriculum standards. Produced by Marble.

Version: v1 · Topics: 1,591 · Prerequisite edges: 3,225 · Subjects: 8

See it

Every dot is a micro-topic, colored by subject; height is age; each thread is a prerequisite (full-quality video). Explore it interactively at withmarble.com/curriculum — tap any concept to trace everything a learner must master before it.

What this is

Most curriculum data is either a flat list of standards or locked inside a product. This dataset is a connected graph of learning:

• 1,591 micro-topics — a single, teachable idea (e.g. “Building sentences”, “Apparent brightness of stars”), each with a plain-language description, mastery evidence criteria, a type (conceptual / procedural / representational / language / meta), a subject + domain, and an approximate age range.
• 3,225 prerequisite dependencies — a directed acyclic graph: “topic X depends on prerequisite Y”, each edge tagged hard/soft and carrying a one-line reason.
• Curriculum alignment — each micro-topic links to the standards it was distilled from (NGSS, Common Core, the UK National Curriculum, and more).
• Teaching research — 1,051 records covering the current Science and Mathematics topics: teaching approaches, concrete learning activities, feedback, assessment guidance and linked sources. Coverage is partial; the other 540 topics have no research in this release.
• Domain clusters — 176 parent-friendly one-paragraph summaries per (subject, domain, age band).

Subjects

Files

All data lives in data/ as UTF-8 JSON. See schema/ for JSON Schemas and manifest.json for counts + SHA-256 checksums.

A topic

{
  "id": "mt_N8CpN1EJrP",
  "type": "CONCEPTUAL",
  "subject": "English",
  "domain": "Grammar & Punctuation",
  "name": "Building sentences",
  "description": "Understand that words combine to make sentences — a sentence expresses a complete thought…",
  "ageRangeStart": 4,
  "ageRangeEnd": 6,
  "centrality": 0.257,
  "evidence": [
    "Distinguish between complete sentences and fragments",
    "Compose a complete sentence with a subject and verb"
  ],
  "assessmentPrompt": "If {{name}} says something like \"The dog\", can they tell you that's not a complete sentence…?",
  "standards": ["ccss-ela:L.K.1f", "uk-nc-2013:Eng.App2.Y1.Sent.1"]
}

• id — stable identifier (mt_…), referenced by dependencies and by neighbours.
• standards — keys into curriculum-standards.json ("<curriculum-slug>:<code>").
• assessmentPrompt — a natural-language check for the idea. Contains a {{name}} placeholder (the child’s name); substitute or strip before display.

A dependency

{ "topicId": "mt__00ZSLnB7p", "prerequisiteId": "mt_VBl1T1sFCM", "strength": "hard",
  "reason": "Must understand vibrations make sound before finding volume patterns" }

topicId depends on prerequisiteId. Reverse the edge to get “unlocks”.

Teaching research

data/teaching-research.json adds how a topic might be taught to the graph of what children learn. It covers 547 Science and 504 Mathematics topics. coverage.missingTopicIds and per-subject counts make the 540 topics without research explicit; absence is not a judgement about whether a topic can be taught.

The file has schemaVersion: 1 and taxonomy version: "v1". Each records[] entry joins topics.json through topicId and contains:

sourceClaims exactly match sources[].claim after publication transformations. A source may support a general principle or curriculum alignment; this does not mean the proposed activity was tested. Some activities have no directly linked source claim.

These records were agent-authored. The learning guidance was rewritten for portability, checked by a separate factual critic, and sampled by a second agent. Most original pedagogy and assessment were retained and may contain claims that the updated guidance qualifies; chart-specific alignment repairs after #30 are recorded in the affected learning-guidance review notes; read the constraints and review notes alongside them. Sources were not re-fetched or re-verified for this release. Historical verified: true and confidence: "high" are not new evidence ratings. Activities are research-informed proposals, not empirically validated interventions.

Some original research quoted restricted curriculum standards. Matched passages are replaced with inline curriculum references, including inside citation claims; all transformed paths are listed in publicationNotes. This is a conservative mechanical omission pass, not comprehensive rights clearance. Consult the linked upstream sources for their wording and terms. See PROVENANCE.md.

After the chart split in #30, mt_bESTSBB0wK covers pie charts, mt_H6LlpWgEYS covers earlier line-graph problem-solving, and mt_HlNSurPVDp has a separate construction-and-interpretation record. The new record adapts existing graph research; it is not a fresh literature review, and all reused source-verification flags are false. Its assessment requires a complete constructed graph. The later grouped-charts assessment routes graph-choice errors to data presentation and sector-angle errors to pie charts.

The file is bound to the exact topics.json bytes through topicFileSha256. Regenerate the research projection when the canonical topic file changes, even when topic IDs remain stable.

Using it

Pure data — no runtime, no dependencies. Load the JSON and go.

import topics from './data/topics.json' with { type: 'json' };
import deps from './data/dependencies.json' with { type: 'json' };

const byId = new Map(topics.topics.map(t => [t.id, t]));
const prereqs = deps.dependencies
  .filter(d => d.topicId === 'mt_N8CpN1EJrP')
  .map(d => byId.get(d.prerequisiteId).name);

Join teaching guidance to a topic:

import research from './data/teaching-research.json' with { type: 'json' };

const guidanceByTopic = new Map(research.records.map(r => [r.topicId, r]));
const guidance = guidanceByTopic.get('mt_N8CpN1EJrP'); // undefined: English is not covered yet
const activities = guidance?.learn.activities ?? [];

Validate structure, research coverage, evidence links and checksums:

node scripts/validate.mjs
node --test scripts/teaching-research.test.mjs

License

This dataset is multi-licensed — read this before you use or redistribute it.

Why share-alike + still commercial-friendly: ODbL distinguishes a derivative database (extend/modify the taxonomy → must stay open) from a produced work (use it inside a product, model, or app → stays yours). So you can build a commercial product on this without open-sourcing your product; you only owe back improvements to the taxonomy itself.

Attribution

Any use must credit:

Marble Skill Taxonomy (v1) · © Generative Spark, Inc. (Marble) · https://withmarble.com · licensed under ODbL 1.0 (database) and CC BY-SA 4.0 (content).

Plus the upstream notices in PROVENANCE.md for any curriculum standards you use. See CITATION.cff for a formal citation.

What’s not here

Deliberately excluded from this release: semantic embeddings (derived, recomputable) and any per-child / user data (never published). See CHANGELOG.md.