简单来说
把一篇著名的机器学习论文从开放获取的源文件重新排版:公式有编号、可以引用,定理和引理放在框里,有证明,参考文献用作者–年份格式。
成品一览
Rafailov、Sharma、Mitchell、Ermon、Manning和Finn的论文Direct Preference Optimization(NeurIPS 2023),从arXiv源文件节选,排成7×10英寸、十页的预印本。梅紫色色带上印着标题,背景是一束逻辑斯谛曲线,即DPO梯度给每个样本加的权重的形状。下面是一栏宽松的Spectral正文:十三个编号公式,第5节的定义、引理和定理排在框里、标签与正文同行,一个证明概要和一个完整证明都以方块结尾,一幅用代码重画的流程图,以及按作者–年份引用的二十种文献。正文中对公式、命题和章节的每一处引用都是链接,每个编号都来自一个标签,和LaTeX一样。
这道食谱解答
- 怎样在数学论文中给公式编号、陈述定理、排证明,并引用它们?
- 怎样排数学公式(行内、行间、编号公式),并在PDF中保持矢量?
- 怎样用Postext重新排版arXiv或PubMed Central上的开放获取论文,并保留引用、图和许可声明?
简短回答
// \label{eq:x} in a display formula numbers it, on its row of an align; a box opened as
// :::callout{type="lemma" #lem:x} counts as a statement. \eqref{eq:x}, \ref{lem:x} and
// :ref{id="lem:x"} print the number and link to it.
const equationNumbering = { // (1) to (13): one sequence through the paper and its appendix
numberingTemplate: '{n}', resetOn: 'never', format: '({n})' };
// A proof: an italic label with no number, and a □ flush right on its last line, drawn
// from TeX's \square because maths runs (Spectral has no □ of its own).
const proofOf = (label) => ({ endMark: '□',
numbering: { label, counter: false, bold: false, italic: true } });
const statements = [ // a counter per kind, as the paper has it: Definition 1, Lemma 1, Theorem 1
{ id: 'definition', numbering: { label: 'Definition' } },
{ id: 'lemma', numbering: { label: 'Lemma' } }, // counter: 'theorem' would share one sequence
{ id: 'theorem', numbering: { label: 'Theorem' } },
{ id: 'proof', ...proofOf('Proof') },
{ id: 'sketch', ...proofOf('Proof Sketch') },
];
用料
做法
#1 · 给公式和命题加标签,按键名引用
代码就是上面的简短回答。从postext 1.19起,引擎自己维护LaTeX的那些计数器。行间公式里的\label{eq:dpo}按阅读顺序给它编号;align里每一行带标签的各得一个编号,标了\notag的行没有;resetOn: 'never'让编号从(1)一直数到附录里的(13),与这篇节选论文重新编排的编号一致。用:::callout{type="lemma" #lem:same-policy}打开的框以“Lemma 2.”开头,因为它的样式带有numbering: { label: 'Lemma' }和独立的计数器。正文里\eqref{eq:dpo}印出(7),\ref{lem:same-policy}印出2,:ref{id="lem:same-policy"}印出Lemma 2,都链接到各自的目标;Eq.~\eqref{…}保留LaTeX的不断行空格。附录里重述的引理以***:ref{id="lem:same-preference"} Restated.***开头,引用沿用周围标签的粗体。章节不需要帮忙:编号来自标题模板,{startAt=3}在删去第2节之后保留论文原来的编号,crossRefs.section像原文一样印出首字母大写的“Section 5”。
#2 · 每种环境一种框
const box = ({ id, ...numbered }, extra) => ({ id, marginTop: pt(LEAD * 0.6),
marginBottom: pt(LEAD * 0.6), backgroundEnabled: false,
snapToGrid: false, // exact space round a statement, as amsthm's \topsep; one column, no grid
padding: { top: mm(1.8), right: mm(4), bottom: mm(1.8), left: mm(4) },
body: { italic: true, firstLineIndent: pt(0), boldColor: col('ink') }, ...extra, ...numbered });
const stripe = (color) => ({ enabled: true, side: 'left', width: pt(3), color: col(color) });
const proof = { keepTogether: false, // a long proof runs on to the next page
// A proof is upright text with an italic run-in label, set off by space alone.
padding: { top: pt(0), right: pt(0), bottom: pt(0), left: pt(0) },
body: { firstLineIndent: pt(0), italicColor: col('ink') } };
const look = { definition: { stripe: stripe('rule') }, lemma: { stripe: stripe('accent') },
theorem: { backgroundEnabled: true, background: col('tint') }, proof, sketch: proof };
const theoremStyles = [...statements.map((s) => box(s, look[s.id])),
box({ id: 'restated' }, look.lemma)]; // Lemma 1 again in the appendix, with no new number
每种环境都是一个标注框样式,只用一种手段:定义用灰色侧边条,引理用梅紫色侧边条,定理用浅梅紫色底。box()把这些外观加在简短回答的编号设置上。body.italic像amsthm的plain样式一样把陈述排成斜体,同行标签在其中保持粗正体。证明不加框:标签为斜体、不带编号(counter: false),endMark: '□'在最后一行右端排一个方块,因为数学引擎在运行,方块用TeX的\square绘制;keepTogether: false让较长的证明可以转到下一页。snapToGrid: false让每个命题上下的空白都相同:单栏没有相邻栏需要对齐,精确的间距比取整到基线网格的更好看。
#3 · 用论文自己的文献表做作者–年份引用
registerCitationEngine(createCiteprocEngine({ styles: STYLES, locales: LOCALES }));
const citations = { style: 'harvard-cite-them-right', link: true,
bibliography: { fontSize: pt(8.2), lineHeight: pt(10.8), hangingIndent: mm(5),
entrySpacing: pt(1.6), doi: 'link' } };
参考文献是:::references{format=bibtex}块里的BibTeX,根据论文的.bbl重建,只保留节选正文仍在引用的二十种文献。Cite Them Right Harvard在正文中排成“(Bong and Rinaldo, 2022)”,在文献表中排成“Ouyang, L. et al. (2022)”,四位及以上作者只列第一位。正文里写@ziegler2020finetuning得到叙述式引用Ziegler et al. (2020)。文献表放在:::bibliography所在的位置,即第7节之后、附录之前,与原论文相同。它前面的:::pagebreak让附录像会议论文那样另起一页;否则A.1那个四行的公式放不下,文献表最后一页的页脚会空出五行。
#4 · 标题放在色带上,来源放在页脚
const text = (id, content, family, size, color, placement, extra) => ({ kind: 'text', id,
content, fontFamily: family, fontSize: pt(size), color: col(color), align: 'left',
overflow: 'wrap', placement, ...extra });
const at = (to, edge, x, y, width) => ({ anchor: { to, edge }, offset: { x: mm(x), y: mm(y) },
...(width && { size: { width: mm(width), height: 'auto' } }) });
const caps = (size, fontWeight = 600) => ({ fontWeight, letterSpacing: pt(size * 0.18),
textTransform: 'uppercase' });
const BAND = 80; // mm from the trim's top to the foot of the band
const titleBlock = {
enabled: true,
minHeight: mm(BAND + 13 - TOP), // the band, then the byline: the abstract starts under it
slot: { elements: [
{ kind: 'box', id: 'band', style: { backgroundColor: col('accent') },
placement: { anchor: { to: 'bleed', edge: 'top-left' },
size: { width: 'fill', height: mm(BAND + 3) } } }, // + 3 mm of bleed
{ kind: 'image', id: 'curves', resourceId: 'sigmoids', // drawn in code: σ(βu)
placement: at('page', 'top-left', 70, 3, 108) },
text('kicker', '{attr.kicker}', SANS, 7.5, 'mist',
at('page', 'top-left', INNER, 30, 120), caps(7.5)),
text('title', '{titleText}', SERIF, 25, 'paper', at('#kicker', 'below', 0, 3.5, MEASURE),
{ fontWeight: 600, lineHeight: 1.08 }),
text('venue', '{attr.venue}', MONO, 7, 'mist',
at('page', 'top-left', INNER, BAND - 7, 120), { overflow: 'clip' }),
text('authors', '{attr.authors}', SANS, 9.5, 'ink', at('page', 'top-left', INNER, BAND + 6,
MEASURE), { fontWeight: 500, lineHeight: 1.45 }),
text('affiliations', '{attr.affiliations}', SANS, 7.5, 'muted',
at('#authors', 'below', 0, 1.4, MEASURE)),
] },
};
const head = (id, content, parity, edge, x, extra) => text(id, content, SANS, 7.5, 'muted',
at('page', `top-${edge}`, x, 13), { overflow: 'clip', ...caps(7.5, 500), align: edge,
parity, pages: 'body', ...extra });
const folio = { fontWeight: 700, color: col('accent') };
// The running heads are the paper's short forms: checks.contentInConfig allows them (#849).
const RUNNING = ['Rafailov, Sharma, Mitchell, Ermon, Manning & Finn',
'Direct Preference Optimization'];
const header = { elements: [
head('v-folio', '{pageNumber}', 'even', 'left', OUTER, folio),
head('v-title', RUNNING[0], 'even', 'left', OUTER + 9),
head('r-title', RUNNING[1], 'odd', 'right', -(OUTER + 9)),
head('r-folio', '{pageNumber}', 'odd', 'right', -OUTER, folio),
] };
const footer = { elements: [ // the first page: where the text comes from, and its licence
text('source', '{attr.source}', MONO, 6.2, // the H1's attribute, as kicker and venue are
'muted', at('page', 'bottom-left', INNER, -13, MEASURE - 10), { pages: 'opener' }),
text('drop-folio', '{pageNumber}', SANS, 7.5, 'accent', at('page', 'bottom-right', -OUTER, -13),
{ ...folio, align: 'right', pages: 'opener', overflow: 'clip' }),
] };
标题是全文唯一的一级标题,它的paper样式改为绘制一组设计元素:色带从页面顶部出血,标题按版心宽度折行,作者和单位来自标题属性,其中的¹、²和*直接写成字符。页脚里带pages: 'opener'的元素只在第1页印出来源、许可和所做的改动,这样第一页单独流传时也能说明文本出处。
#5 · 用代码重画插图
const R = (x) => Math.round(x * 100) / 100;
const svg = (w, h, body) => `<svg xmlns="http://www.w3.org/2000/svg" width="${w * PX_PER_MM}" `
+ `height="${h * PX_PER_MM}" viewBox="0 0 ${w} ${h}">${body}</svg>`;
function sigmoids() { // 108 × 24 mm over the kicker: σ(βu) for β from 0.25 to 4
let out = '';
[0.25, 0.4, 0.6, 1, 1.6, 2.5, 4].forEach((beta, i) => {
const pts = Array.from({ length: 105 }, (_, k) => {
const u = (k - 52) / 8; // u from −6.5 to 6.5
return `${k ? 'L' : 'M'}${R(2 + k)} ${R(22 - 20 / (1 + Math.exp(-beta * u)))}`;
}).join('');
out += `<path d="${pts}" fill="none" stroke="${palette.mist}" stroke-width="${R(0.3
+ i * 0.06)}" stroke-opacity="${R(0.25 + i * 0.1)}"/>`;
});
return svg(108, 24, out);
}
function tex(markup, x, y, size, color = 'ink', anchor = 0) { // maths as MathJax paths
const r = renderMath(markup, false, 100);
const k = size / 1000;
return `<g transform="translate(${R(x - anchor * r.viewBox.width * k)} ${R(y)}) scale(${k})" `
+ `fill="${palette[color]}">${r.paths.map((p) => `<path d="${p.d}"/>`).join('')}</g>`;
}
// Each label names its family and weight: loadSvg embeds the Work Sans files they set.
const words = (x, y, s, { size = 2.6, weight = 400, color = 'ink', anchor = 'start' } = {}) =>
`<text x="${R(x)}" y="${R(y)}" font-family="${SANS}" font-weight="${weight}" `
+ `font-size="${size}" fill="${palette[color]}" text-anchor="${anchor}">${s}</text>`;
const rect = (x, y, w, h, stroke, fill = 'paper', r = 1.2) => `<rect x="${x}" y="${y}" `
+ `width="${w}" height="${h}" rx="${r}" fill="${palette[fill]}" stroke="${palette[stroke]}" `
+ 'stroke-width="0.35"/>';
function arrow(points, color) { // a line with a drawn head, never an SVG marker
const [[x1, y1], [x2, y2]] = points.slice(-2);
const a = Math.atan2(y2 - y1, x2 - x1);
const head = [a - 0.45, a + 0.45].map((t) => `${R(x2 - 2 * Math.cos(t))} ${R(y2 - 2
* Math.sin(t))}`);
return `<path d="M${points.map(([x, y]) => `${R(x)} ${R(y)}`).join('L')}" fill="none" `
+ `stroke="${palette[color]}" stroke-width="0.45"/><path d="M${head[0]}L${R(x2)} ${R(y2)}`
+ `L${head[1]}Z" fill="${palette[color]}"/>`;
}
function preferences(x, y, w) { // a prompt and two answers, the preferred one first
const page = (dx) => rect(x + dx, y + 15.5, 6, 7.2, 'rule', 'paper', 0.6)
+ [0, 1, 2].map((l) => `<path d="M${x + dx + 1.2} ${y + 17.6 + l * 1.6}h3.6" `
+ `stroke="${palette.rule}" stroke-width="0.4"/>`).join('');
return rect(x, y, w, 25, 'rule', 'tint')
+ words(x + 2.4, y + 4.6, 'PREFERENCE DATA', { size: 2.1, weight: 600, color: 'muted' })
+ tex('x', x + 2.4, y + 9.8, 3) + words(x + 4.3, y + 9.8, ': “write me a poem about',
{ size: 2.25 }) + words(x + 5.2, y + 13, 'the history of jazz”', { size: 2.25 })
+ page(2.4) + tex('y_w', x + 9.4, y + 20.6, 3) + tex('\\succ', x + 14, y + 20.4, 3, 'accent')
+ page(18.4) + tex('y_l', x + 25.4, y + 20.6, 3);
}
function pipeline() { // 130 × 64 mm: RLHF on the left, DPO on the right
const title = (x, name, sub, color) => words(x, 4, name, { size: 3.2, weight: 600, color })
+ words(x, 8, sub, { size: 2.3, color: 'muted' });
const model = (x, y, w, label, color) => rect(x, y, w, 9, color)
+ words(x + w / 2, y + 5.8, label, { size: 2.6, weight: 600, color, anchor: 'middle' });
const note = (x, y, lines, color, anchor) => lines.map((line, i) => words(x, y + i * 3, line,
{ size: 2.3, color, anchor })).join('');
return svg(MEASURE, 64, title(0, 'RLHF', 'Reinforcement learning from human feedback', 'ochre')
+ preferences(0, 14, 35) // 35 mm: the prompt's first line ends 32.6 mm into the box
+ arrow([[36, 26.5], [48, 26.5]], 'ochre') + note(42, 22.4, ['maximum'], 'ochre', 'middle')
+ note(42, 31.4, ['likelihood'], 'ochre', 'middle')
+ model(49, 22, 25, 'reward model', 'ochre') + model(49, 50, 25, 'LM policy', 'ochre')
+ arrow([[55, 31.5], [55, 49]], 'ochre')
+ note(53.4, 41.6, ['label rewards,', 'reinforcement', 'learning'], 'ochre', 'end')
+ arrow([[68, 49], [68, 31.5]], 'ochre') + note(69.6, 39, ['sample', 'completions'], 'ochre')
+ `<path d="M88 2V62" stroke="${palette.rule}" stroke-width="0.3"/>`
+ title(93, 'DPO', 'Direct preference optimization', 'accent')
+ preferences(93, 14, 37)
+ arrow([[111.5, 39.5], [111.5, 49]], 'accent')
+ note(113.2, 43.4, ['maximum', 'likelihood'], 'accent', 'start')
+ model(99, 50, 25, 'final LM', 'accent'));
}
图1按原图的方框和文字画成SVG。标签在font-family里写明Work Sans,loadSvg把SVG交给引擎的inlineSvgFonts,由它嵌入标签用到的两个Work Sans文件;tex()把y_w ≻ y_l写成MathJax路径,与公式用的是同一套字形。箭头是画出来的路径,不用SVG的marker,所以图在PDF里仍是矢量。
完整食谱
// ═══ Postext Cookbook · Nº 138 · Machine-learning paper with theorems and proofs ═══
// https://postext.dev/en/cookbook/ml-paper-theorems-proofs
// Code: MIT · Text: Rafailov et al. 2023, arXiv:2305.18290 (CC BY 4.0), abridged · Art: code
// Fonts: Spectral, Work Sans, JetBrains Mono (SIL OFL 1.1) · Needs postext ≥ 1.65.0
// DPO (NeurIPS 2023) re-set as a preprint: numbered equations with labels and references,
// definition, lemma and theorem boxes, proofs that end in a square, author–year citations.
import {
buildDocumentWithFonts, renderPageToCanvas, registerResourceImage, registerCitationEngine,
defaultResourceTypes, initMathEngine, renderMath, inlineSvgFonts,
} from 'https://esm.sh/postext';
import { renderToPdf, decompressWoff2 } from 'https://esm.sh/postext-pdf';
import { createCiteprocEngine, STYLES, LOCALES } from 'https://esm.sh/postext-citeproc';
const LANG = 'en'; // @lang: the language of the sample document ('en' | 'es')
const RECIPE = 'ml-paper-theorems-proofs';
// ─── 1 · Design ─────────────────────────────────────────────────────────────
const palette = {
ink: '#1d1a22', accent: '#6b2a5f', // text; plum: the band, labels, theorem stripes
ochre: '#a8621c', tint: '#f4edf2', // the RLHF route in Figure 1; the theorem fill
rule: '#cfc3cc', muted: '#675f6b', // hairlines; running heads, notes
mist: '#e3c9dc', paper: '#ffffff', // type and curves on the band; white
};
const col = (id) => ({ hex: palette[id], model: 'hex', paletteId: id });
const colorPalette = Object.entries({ ...palette, 'main-color': palette.accent })
.map(([id, hex]) => ({ id, name: id, value: { hex, model: 'hex' } }));
const [SERIF, SANS, MONO] = ['Spectral', 'Work Sans', 'JetBrains Mono'];
const [BODY, LEAD] = [9.6, 13.4]; // pt: one column of 130 mm, about 74 characters a line
// mm: a 7 × 10 in trim, as technical books and many preprint series print
const [TRIM_W, TRIM_H, TOP, BOTTOM, INNER, OUTER] = [178, 254, 22, 22, 21, 27];
const MEASURE = TRIM_W - INNER - OUTER;
// #region answer: amsthm in Markdown: numbered equations, theorems and proofs
// \label{eq:x} in a display formula numbers it, on its row of an align; a box opened as
// :::callout{type="lemma" #lem:x} counts as a statement. \eqref{eq:x}, \ref{lem:x} and
// :ref{id="lem:x"} print the number and link to it.
const equationNumbering = { // (1) to (13): one sequence through the paper and its appendix
numberingTemplate: '{n}', resetOn: 'never', format: '({n})' };
// A proof: an italic label with no number, and a □ flush right on its last line, drawn
// from TeX's \square because maths runs (Spectral has no □ of its own).
const proofOf = (label) => ({ endMark: '□',
numbering: { label, counter: false, bold: false, italic: true } });
const statements = [ // a counter per kind, as the paper has it: Definition 1, Lemma 1, Theorem 1
{ id: 'definition', numbering: { label: 'Definition' } },
{ id: 'lemma', numbering: { label: 'Lemma' } }, // counter: 'theorem' would share one sequence
{ id: 'theorem', numbering: { label: 'Theorem' } },
{ id: 'proof', ...proofOf('Proof') },
{ id: 'sketch', ...proofOf('Proof Sketch') },
组合好的script.js可以直接运行:把它粘贴到任何页面的模块脚本中,或在CodePen上打开这道食谱。 GitHub上的食谱文件夹 ↗ (在新标签页中打开)
变化
#像NeurIPS模板那样用数字引用
原论文用natbib的数字引用;IEEE在正文中印[1],文献表按引用顺序编号。
-const citations = { style: 'harvard-cite-them-right', link: true,
+const citations = { style: 'ieee', link: true,#引理和定理统一编号
有些期刊把所有命题排成一个序列(引理1、引理2、定理3),就像\newtheorem{lemma}[theorem]那样。让引理使用定理的计数器。
- { id: 'lemma', numbering: { label: 'Lemma' } }, // counter: 'theorem' would share one sequence
+ { id: 'lemma', numbering: { label: 'Lemma', counter: 'theorem' } },常见问题
易错点
数学公式需要在首次构建前await initMathEngine()
在数学引擎运行之前排版的公式是灰色的占位方框。从https://esm.sh/postext连同其他符号一起导入initMathEngine,并在首次构建前await它;含有数学公式的文档不经这一步就构建时,控制台会打印一条警告。从postext 1.5起,MathJax作为一个模块随包发布,由initMathEngine去获取,所以从CDN直接导入就能用,不需要?bundle(1.4及以前,所有符号都必须来自https://esm.sh/postext?bundle)。只从一个URL导入:两个URL会加载两个引擎。 数学 →
- 只有数学引擎运行时,
endMark才用TeX绘制方块(□ ∎ ■)。没有initMathEngine()的文档里,结束符用框内正文的字体排,而Spectral和Fontsource所有latin文件一样没有□:这时请选一个字体能排的结束符。 - 各栏齐底会在标题上方加基线网格行,把被高公式留短的页面填满。
balancing: { maxLinesPerHeading: 1 }让论文的标题基本保持原有间距,这样的页面可能会提前几行结束。 \underbrace的说明文字比括号还宽时(如第4节的梯度公式),需要用aligned拆成两行;比版心宽的公式会向右超出版心。
致谢
- 文本
- Abridged from Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C. D. and Finn, C. (2023) “Direct Preference Optimization: Your Language Model is Secretly a Reward Model”, Advances in Neural Information Processing Systems 36 (NeurIPS 2023), arXiv:2305.18290v3. Sections 2 and 5.2, most of 6 and appendices A.2–A.4, A.6 and B–E cut, other paragraphs and sentences shortened; section numbers kept, equations renumbered in the abridged order; citations re-set author–year; Figure 1 redrawn · Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, Chelsea Finn · CC BY 4.0
- 字体
- Spectral (SIL OFL 1.1) · Work Sans (SIL OFL 1.1) · JetBrains Mono (SIL OFL 1.1)


