Formats and versions

Markdown, JSON and plain text for every paper, figure assets, versioned and unversioned URLs, and how responses are cached.

Every format is a rendering of the same document tree.

Formats

Extension	Content-Type	Contents
`.md`	`text/markdown; charset=utf-8`	YAML front matter, `# Title`, `## Abstract`, sections, LaTeX math, GFM tables, figures, `## References`, footnotes
`.json`	`application/json`	Metadata, build info, token count, the full tree and inferred fields. See JSON schema
`.txt`	`text/plain; charset=utf-8`	Plain text: headings as plain lines, math kept as LaTeX, tables tab-separated, `References` and `Footnotes` at the end

curl https://arxivv.org/1706.03762.md
curl https://arxivv.org/1706.03762.json
curl https://arxivv.org/1706.03762.txt

Markdown conventions

Element	Rendering
Front matter	`arxiv`, `version`, `title`, `authors`, `date`, `categories`, `license` (when known), `url` (arXiv abstract)
Title	`# Attention Is All You Need`
Abstract	`## Abstract`
Section heading	Level-1 sections are `##`, subsections `###`, then `####`: `### 3.2 Attention {#S3.SS2}`
Inline math	$d_k$
Display math	`$$...$$` followed by `(1)` on the same line when numbered
Citation	`[12]`, as printed in the paper
Cross-reference	`[2](#S3.F2)`: the printed text, linked to the target id
Figure	`![alt](https://arxivv.org/1706.03762v7/assets/Figures/ModalNet-21.png)` then `Figure 1: caption`
Table	`Table 1: caption` then a GFM table (first row is the header row, `\|` escaped)
Theorem	`Theorem 1 (title).` then the body; proofs start `Proof.`
Footnote	`[^footnote3]`, defined at the very end
Bibliography	Numbered list under `## References`

Image URLs are absolute and point at the asset route. Alt text is the image’s own alt, or the figure caption.

---
arxiv: "1706.03762"
version: 7
title: "Attention Is All You Need"
authors: "Ashish Vaswani, Noam Shazeer, Niki Parmar, ..."
date: "2023-08-02"
categories: ["cs.CL", "cs.LG"]
license: "http://creativecommons.org/licenses/by/4.0/"
url: "https://arxiv.org/abs/1706.03762v7"
---

# Attention Is All You Need

## Abstract

The dominant sequence transduction models are based on complex recurrent or
convolutional neural networks...

## 3 Model Architecture {#S3}

### 3.2 Attention {#S3.SS2}

/<paper>/sections/<anchor>.md returns one section subtree in the same form, heading included, followed by the definitions of any footnotes it uses.

Figure assets

/<paper>/assets/<path> serves a file referenced by arXiv’s HTML, such as a figure image. <path> is the image’s src in the tree, e.g. Figures/ModalNet-21.png.

Property	Value
Availability	Mirrored from arXiv asynchronously after the first build; `404 not-found` until mirrored
Content-Type	As stored from arXiv, else `application/octet-stream`
Cache-Control	Versioned: `public, max-age=31536000, immutable`; unversioned: `public, max-age=3600, stale-while-revalidate=86400`
Headers	`ETag`, `X-Arxivv-Schema`, `X-Arxivv-Version`; `Cache-Tag: paper:<id>` on unversioned requests

curl -O https://arxivv.org/1706.03762v7/assets/Figures/ModalNet-21.png

Versions

URL	Serves	Extra headers
`/1706.03762.md`	Latest version	`Link: </1706.03762v7.md>; rel="canonical"`, `Cache-Tag: paper:1706.03762`
`/1706.03762v7.md`	Version 7, forever	None
`/1706.03762v1.md`	Version 1, forever	None

Unversioned URLs follow arXiv: they serve the latest known version.
Every response states its version in X-Arxivv-Version.
Versioned URLs are immutable. Cite them.
An unknown paper or version is looked up on arXiv (OAI-PMH) on first request.
Old-style IDs work the same way: /hep-th/9901001.md, /hep-th/9901001v1.json.

Daily harvest

Property	Value
Schedule	Daily at 03:00 UTC
Source	arXiv OAI-PMH `ListRecords` (`arXivRaw`) for the previous day
Effect	Updates latest-version pointers and metadata
Purge	When a known paper gains a version, unversioned responses tagged `paper:<id>` are purged from cache
Builds	The new version is built on its first request, not by the harvest

Caching

Response	`Cache-Control`
Versioned URL	`public, max-age=31536000, immutable`
Unversioned URL	`public, max-age=3600, stale-while-revalidate=86400`
`.json`, `outline.json` or `metadata.json` while inferred fields are `pending` or `failed`	`public, max-age=30`
`metadata.json` before the tree is built	`public, max-age=30`
A `Prefer` preference was applied	`no-store`
Errors	`no-store`

Assets follow the URL rule: versioned asset URLs are immutable, unversioned ones (/1706.03762/assets/...) use the unversioned policy and carry Cache-Tag.

/<paper>/metadata.json never triggers a build and ignores Prefer, so it is never no-store. Once the tree is built and inferred fields are ready, it follows the URL rule like any other route.

Precedence: no-store first, then pending or failed, then the URL rule. Every paper response carries Vary: Prefer.

Response headers

Every arXivv paper response carries these headers:

Header	Example	Meaning
`X-Arxivv-Schema`	`1`	Major schema version
`X-Arxivv-Version`	`7`	arXiv version served
`X-Arxivv-Quality`	`0.97`	Conversion quality score
`X-Arxivv-Source`	`arxiv-html`	Where the tree was built from
`Vary`	`Prefer`	Responses differ by `Prefer`
`Cache-Control`	See caching
`Link`	`</1706.03762v7.md>; rel="canonical"`	Unversioned requests only
`Cache-Tag`	`paper:1706.03762`	Unversioned requests only (assets too); used to purge on new versions

X-Arxivv-Quality and X-Arxivv-Source are absent from 202 responses and from metadata.json for an unbuilt paper, since no tree exists yet. See first request for Preference-Applied, Retry-After and Location.