Academic PDF To Gfm

Academic PDF → GitHub GFM Conversion

A battle-tested workflow for converting academic/research PDF papers into GitHub-renderable GFM markdown with inline figures, mathematically correct LaTeX, and validated output.

Battle-tested on: López de Prado (2026) "How to Use the Sharpe Ratio" — 51 pages, 82 equations, 8 figures.

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Quick Start — Per-Page Routing Workflow

Do NOT try to apply a single extraction method to the entire PDF. Modern academic papers mix text, tables, and figures; route each page individually.

Step 1: Render all pages at 300 DPI

# Generate PNG for every page at 300 DPI
uv run --python 3.14 --with pymupdf python3 << 'EOF'
import fitz, os
doc = fitz.open('paper.pdf')
os.makedirs("pages", exist_ok=True)
for i in range(len(doc)):
    pix = doc[i].get_pixmap(matrix=fitz.Matrix(300/72, 300/72), alpha=False)
    pix.save(f"pages/page{i:04d}.png")
    print(f"Page {i}: {pix.width}×{pix.height}")
EOF

Step 2: Detect page type for each page

# Locate the Unlimited-OCR script. If installed via marketplace, it will be in the cc-skills plugin directory.
# If running from the cc-skills repo directly:
S=${S:-~/eon/cc-skills/plugins/unlimited-ocr/scripts/unlimited_ocr.py}

# For each page, run Unlimited-OCR and inspect the <|det|> markers
for img in pages/page*.png; do
    echo "=== $(basename $img) ==="
    uv run --no-project $S parse --input "$img" --collapse-math-spacing --quiet 2>&1 | \
        grep "<|det|>" | cut -d'>' -f2 | cut -d' ' -f1 | sort | uniq -c
done

Step 3: Route each page

Based on the detected <|det|> markers:

Step 4: Assemble and validate

# Concatenate all page outputs
cat pages/page*.md > paper-raw.md

# Extract embedded images (optional)
uv run --python 3.14 --with pymupdf python3 references/extract-images.py paper.pdf

# Validate math before pushing
node references/validate-math.mjs paper-raw.md --fix

Tool Comparison (by page type)

A single academic PDF is routinely born-digital prose on most pages and equation-dense or figure-dense on a few. Routing the entire document as one type wastes capability and silently loses content.

The Per-Page Routing Strategy

For every page in the PDF:

1. Render at 300 DPI to a PNG (required for Unlimited-OCR accuracy; 150 DPI hallucinates silently)
2. Detect page type by running Unlimited-OCR locally (~2.4 s/page on Apple Silicon)
3. Route based on what the model found:
   • Text-only pages (no math, no charts): extract with pymupdf4llm or pdftotext (cheaper, preserves structure)
   • Equation-bearing pages: use Unlimited-OCR output (returns real LaTeX, e.g. \left\{ \begin{array} ... \end{array} \right\})
   • Chart-bearing pages: segment with Unlimited-OCR, hand each panel to a vision model for description (verified 2026-07-31: charts return empty text from Unlimited-OCR)
   • Fallback: if a page's text-layer extraction is too thin, re-run through Unlimited-OCR

Why Per-Page Matters

The silent failure. Unlimited-OCR localizes charts with perfect bounding boxes and returns zero characters inside. A user who routes the entire document as "Type C, use Unlimited-OCR" on chart-heavy input loses all chart content AND GETS A WELL-FORMED RESULT, so nothing looks wrong until the output is reviewed.

Verified (2026-07-31): Unlimited-OCR detects charts (via <|det|>chart markers) and returns perfect bounding boxes but zero text inside. Pages containing only charts (no text, equations, or tables) would indeed render empty; per-page routing catches this and routes those pages to text-layer extraction or manual inspection instead.

Page Type Detection (Measured at 300 DPI)

Run Unlimited-OCR on every page and inspect the <|det|> markers:

Example: TimeMixer (ICLR 2024, 28 pages)

Verified 2026-07-31 on /Users/terryli/eon/quantml/data/papers/a599/b338e44af70d8e9c87be3c5417bde7864b2c92074e1346703f3e2b641e3d.pdf:

Page 0 — Title page with abstract (route to text extraction)

# Render at 300 DPI
uv run --python 3.14 --with pymupdf python3 -c "
import fitz
doc = fitz.open('timemixer-iclr-2024.pdf')
pix = doc[0].get_pixmap(matrix=fitz.Matrix(300/72, 300/72), alpha=False)
pix.save('page0.png')
"

# Detect structure
S=~/eon/cc-skills/plugins/unlimited-ocr/scripts/unlimited_ocr.py
uv run --no-project $S parse --input page0.png --collapse-math-spacing --quiet 2>&1 | head -15

Output (actual):

<|det|>header [173, 33, 482, 49]<|/det|>Published as a conference paper at ICLR 2024
<|det|>title [172, 99, 818, 149]<|/det|>TIMEMIXER: DECOMPOSABLE MULTISCALE MIXING FOR TIME SERIES FORECASTING
<|det|>text [181, 170, 777, 200]<|/det|>Shiyu Wang \( ^{1,*} \) , Haixu Wu \( ^{2,*} \) , Xiaoming Shi \( ^{1} \) , Tengge Hu \( ^{2} \) , Huakun Luo \( ^{2} \) , Lintao Ma \( ^{1✉} \) , James Y. Zhang \( ^{1} \) , Jun Zhou \( ^{1✉} \)
<|det|>text [182, 200, 781, 243]<|/det|>\( ^{1} \) Ant Group, Hangzhou, China  \( ^{2} \) Tsinghua University, Beijing, China {weiming.wsy,lintao.mlt,peter.sxm,james.z,jun.zhoujun}@antgroup.com, {wuhx23,htg21,luhk19}@mails.tsinghua.edu.cn
<|det|>title [452, 279, 547, 293]<|/det|>ABSTRACT
<|det|>text [230, 311, 770, 575]<|/det|>Time series forecasting is widely used in extensive applications, such as traffic planning…

Routing decision: Only header, title, text regions detected. No equation, no chart → route to pymupdf4llm or pdftotext (faster, preserves structure).

Page 6 — Benchmark table (route to Unlimited-OCR)

uv run --no-project $S parse --input page6.png --collapse-math-spacing --quiet 2>&1 | head -20

Output (actual):

<|det|>text [187, 114, 812, 129]<|/det|>Table 1: Summary of benchmarks. Forecastability is one minus the entropy of Fourier domain.
<|det|>table [175, 131, 825, 287]<|/det|>| Tasks | Dataset | Variate | Predict Length | Frequency | Forecastability | Information |
| Long-term forecasting | ETT (4 subsets) | 7 | 96~720 | 15 mins | 0.46 | Temperature |
| Weather | 21 | 96~720 | 10 mins | 0.75 | Weather |
| Solar-Energy | 137 | 96~720 | 10min | 0.33 | Electricity |
…

Routing decision: table detected → keep Unlimited-OCR output (already in pipe-markdown format, --table-format pipe converts any HTML to markdown).

Page 7 — Figure and table mixed (route Unlimited-OCR for table only)

uv run --no-project $S parse --input page7.png --collapse-math-spacing --quiet 2>&1 | head -30

Output (actual):

<|det|>table [186, 174, 819, 384]<|/det|>| Case | Decompose | Past mixing | Future mixing | M4 | PEMS04 | ETTm1 |
…

Note: Page 7 contains 6 embedded image objects in the PDF (verified via PyMuPDF), but Unlimited-OCR's 300-DPI screenshot rendering does not detect image or image_caption regions. Extract images separately via references/extract-images.py and route them to a vision model for description.

Routing decision: table detected → keep Unlimited-OCR output; extract embedded images separately and hand to a vision model for description (see unlimited-ocr-segment-figure skill).

Type Classification (Historical, kept for reference)

Type A — Word-Generated PDF (Type A: most modern academic papers)

Signs: Embedded fonts, copyable text, Unicode math chars when you copy-paste (∑, π, α, β, γ, →)

Math encoding: Math is Unicode text in PDF stream — NOT images, NOT glyph maps

Consequence: OCR tools like marker-pdf cannot extract LaTeX — they see text like "γ₄" not \gamma_4. They may return empty output or crash silently.

Type B — LaTeX-Generated PDF

Signs: Computer Modern fonts, precise mathematical spacing, arxiv.org source available

Math encoding: Glyph-mapped — structure is partially extractable

Type C — Scanned/Image PDF

Signs: All pages are raster images, zero copyable text

Table Format Notice

Tables from Unlimited-OCR come back as HTML <table> markup (measured by Baidu at ~88% of real tables). The CLI converts them to pipe-markdown by default (--table-format pipe), which is suitable for markdown documents. If you are comparing outputs across readers, pipe-markdown is the canonical form.

Tool Comparison

Recommendation: For per-page routing:

• Text-only pages → pymupdf4llm (preserves formatting) or pdftotext (speed)
• Tables/math pages → Unlimited-OCR (returns real LaTeX and pipe-markdown tables)
• Chart-bearing pages → Unlimited-OCR (detects charts but returns empty text; route to segment-figure skill + vision model for descriptions)
• Mixed pages → Unlimited-OCR (handles tables and layout detection in one pass; segment images separately)

Image Extraction

Save references/extract-images.py:

import fitz, os, sys

doc = fitz.open(sys.argv[1])
os.makedirs("references/media", exist_ok=True)
saved = []
for page_num in range(len(doc)):
    for img_idx, img in enumerate(doc[page_num].get_images(full=True)):
        xref = img[0]
        base_image = doc.extract_image(xref)
        img_bytes = base_image["image"]
        if len(img_bytes) < 2048:   # skip icons/logos/watermarks/rules
            continue
        ext = base_image["ext"]
        fname = f"fig-p{page_num+1:02d}-{img_idx+1:02d}.{ext}"
        with open(f"references/media/{fname}", "wb") as f:
            f.write(img_bytes)
        saved.append((page_num+1, fname, base_image.get("width"), base_image.get("height")))
        print(f"Saved: {fname} ({len(img_bytes)//1024}KB, {base_image.get('width')}×{base_image.get('height')})")
doc.close()
print(f"\n{len(saved)} images saved to references/media/")

Naming: fig-p{page:02d}-{idx:02d}.{ext} — page number in name for easy location matching.

Size filter: Skip < 2 KB (captures icons, watermarks, horizontal rules). Review everything ≥ 2 KB — some are decorative but most are figures.

Insert in markdown:

![Figure 1: Variance of Sharpe ratio estimates](./media/fig-p12-01.png)

Place immediately after the nearest section heading or the paragraph that references the figure.

GitHub GFM Math Rendering Rules

The $$ vs ```math ``` Decision — Root Cause

GitHub's Markdown pre-processor runs BEFORE the math renderer. It treats \\ as an escaped backslash and collapses it to \. This breaks LaTeX line breaks in display math.

The rule is simple:

# BROKEN on GitHub — \\ stripped by pre-processor:

$$
\begin{aligned}
a &= b + c \\
d &= e + f
\end{aligned}
$$

# CORRECT on GitHub:

```math
\begin{aligned}
a &= b + c \\
d &= e + f
\end{aligned}
```

### Display Block Formatting Rules

- `$$` must be on its **own line** — not `$$formula$$` on one line
- **Blank line required** before AND after every `$$` block
- **Blank line required between consecutive** `$$` blocks
- These rules do NOT apply to `` ```math ``` `` blocks

### Supported/Unsupported LaTeX

See [references/github-math-support-table.md](./references/github-math-support-table.md) for the full table.

**Key things to avoid**:

| Command | Problem | Fix |
|---------|---------|-----|
| `\begin{align}` | ❌ Not supported by GitHub | Use `\begin{aligned}` |
| `\boxed{}` | ⚠️ Can cause raw LaTeX passthrough | Remove or use bold text |
| `\operatorname{}` | ⚠️ Active GitHub bug, inconsistent | Use `\text{}` or `\mathrm{}` |
| `\newcommand` | ❌ Was briefly available, then pulled | Expand all macros inline |
| `x^_y` | Superscript immediately before subscript | Write `x^{*}_{i}` with braces |

### Common Gotchas

- `\\[8pt]` vertical spacing inside `$$` → eaten by pre-processor → move to `` ```math ``` ``
- `\frac{1}{T}:\left(` → spurious colon after fraction → remove colon
- Pearson vs excess kurtosis: most finance formulas need Pearson (γ₄ = 3 for Gaussian), not excess. **Always document the kurtosis convention in the formula comment.**
- `\begin{pmatrix}` with `\\` → must use `` ```math ``` ``
- `\begin{cases}` with multiple rows → must use `` ```math ``` ``

---

## GitLab: No Workarounds Needed

**Empirically verified 2026-03-15** on GitLab CE 18.9.2. Confirmed by Comrak source code analysis.

GitLab uses the **Comrak** Rust parser with `math_dollars: true`. When Comrak encounters `$$`, it calls `handle_dollars` which slices the raw input buffer directly and stores it as a `NodeMath` AST node — CommonMark's backslash handler is never invoked on math content. The raw LaTeX is passed to KaTeX via `<span data-math-style="display/inline">` unchanged.

**Every GitHub workaround is unnecessary on GitLab:**

| GitHub problem | GitHub fix required | GitLab |
|----------------|---------------------|--------|
| `\\` in `$$` stripped → broken multiline | Use ` ```math ``` ` | `$$` works with `\\` |
| `\left\{` → `\left{` (delimiter error) | Use `\left\lbrace` | `\left\{` works |
| `\{...\}` set notation → invisible braces | Use `\lbrace...\rbrace` | `\{...\}` works |
| `\,` in `$$` → literal comma | Remove `\,` | `\,` works |
| `\,` in inline `$` → literal comma | Remove `\,` | `\,` works |

On GitLab you can write standard LaTeX without any platform-specific workarounds. If you're targeting GitLab (or hosting your own GitLab CE), skip all the `\lbrace`/`\rbrace` substitutions and ` ```math ``` ` conversions — plain `$$` with standard LaTeX is correct.

### GitLab.com Has a Hard 50-Span Per-Page Limit

**GitLab.com (SaaS) enforces a limit of 50 total math spans per page** (display + inline combined). After the 50th span, all subsequent equations silently fall back to raw LaTeX text. This limit exists to prevent DoS attacks and cannot be overridden on GitLab.com.

| Document math density | gitlab.com | Self-hosted CE |
|---|---|---|
| ≤ 50 total spans | ✅ Renders fully | ✅ |
| 51–100 spans | ⚠️ Partial render | ✅ |
| 100+ spans (academic papers) | ❌ Most equations raw text | ✅ Disable with `math_rendering_limits_enabled: false` |

**Validated on**: Sharpe ratio paper (341 spans) — breaks at span 51 on gitlab.com, renders fully on local CE.

**The W6 check in `validate-math.mjs`** warns when a file exceeds the limit.

**Summary: which platform to use**:
- **GitHub.com**: No math span limit. Use `\lbrace`/`\rbrace` workarounds (handled by `--fix`).
- **Self-hosted GitLab CE**: No limit (disable math_rendering_limits_enabled). No workarounds needed.
- **GitLab.com**: Only suitable for documents with ≤ 50 math spans.

### Self-hosting GitLab CE for Math-Heavy Documents

GitLab CE is free and runs on a single machine. On a 61 GB workstation with slim config:
- Memory footprint: ~3 GB (`puma['worker_processes'] = 2`, `sidekiq['concurrency'] = 5`, monitoring disabled)
- Push mirroring to GitHub: free on CE (syncs within 5 min)
- `glab` CLI: first-party, comparable to `gh`

```yaml
# docker-compose.yml — slim GitLab CE
services:
  gitlab:
    image: gitlab/gitlab-ce:latest
    restart: unless-stopped
    environment:
      GITLAB_OMNIBUS_CONFIG: |
        external_url 'http://YOUR_IP:8929'
        puma['worker_processes'] = 2
        sidekiq['concurrency'] = 5
        prometheus_monitoring['enable'] = false
        alertmanager['enable'] = false
        node_exporter['enable'] = false
        redis_exporter['enable'] = false
        postgres_exporter['enable'] = false
        gitlab_exporter['enable'] = false
    ports: ["8929:8929", "8922:22"]
    volumes:
      - /srv/gitlab/config:/etc/gitlab
      - /srv/gitlab/logs:/var/log/gitlab
      - /srv/gitlab/data:/var/opt/gitlab
```

---

## Validation Pipeline

### Step 1: Install KaTeX Validator

```bash
bun add -g katex   # Bun-first per project policy
# or: npm install -g katex

Step 2: Run Before Every Push

# Validate only (exit 1 on errors)
node references/validate-math.mjs your-file.md

# Validate + auto-fix correctable issues
node references/validate-math.mjs your-file.md --fix

The script is at references/validate-math.mjs. It runs two layers:

Layer 1 — KaTeX syntax: parse errors in $, $$, ```math ``` blocks Layer 2 — GFM structural (issues KaTeX passes but GitHub breaks):

E0 is the most dangerous: a single failing $$ block exposes its $$ delimiters as literal text, creating an orphaned $ that shifts ALL subsequent inline $...$ pairings. One broken equation takes down the entire document.

\{/\} trap: In $$ blocks, \left\{ becomes \left{ (invalid KaTeX delimiter → "Missing or unrecognized delimiter") and \{...\} set notation becomes invisible grouping. Fix: use \lbrace/\rbrace (letter-based, CommonMark-immune). This affects every equation using set notation like \{\hat{SR}_k\} or \min_T\left\{...\right\}.

Exits code 1 on errors (CI-friendly). Warnings do not block CI but should be reviewed.

Local Preview Tools

# GitHub-accurate hot-reload preview
bun add -g @hyrious/gfm
gfm your-file.md --serve

# Offline binary (gh extension)
gh extension install thiagokokada/gh-gfm-preview
gh gfm-preview your-file.md

VS Code extensions:

• shd101wyy.markdown-preview-enhanced — closest to GitHub rendering
• bierner.markdown-preview-github-styles — GitHub CSS styling

Multi-Agent Adversarial Equation Validation

For papers with 10+ equations, use this multi-agent pattern:

Phase 1 — Parallel Extraction

• Agent A: Extract prose with pymupdf4llm, transcribe math from PDF screenshots
• Agent B: Extract and categorize all images

Phase 2 — Parallel Validation

• Agent C: Validate equations against reference implementation (if code/repo exists)
• Agent D: Numerical spot-checks — compute paper's exhibit values, compare

Phase 3 — Discrepancy Handling

• For each discrepancy: write /tmp/paper-discrepancy/eq-{N}.md
• Spawn resolver agents to search online for authoritative third-party sources
• Authority rule: Paper is tentatively more authoritative than code implementation; a third independent source breaks ties

Phase 4 — Guarded Application

• Apply only HIGH-confidence fixes to the markdown
• For MEDIUM-confidence: spawn an independent audit agent before touching the file
• Document all discrepancies even if not fixed — future readers need to know

Anti-Patterns

References

Related Skills

Post-Execution Reflection

After this skill completes, reflect before closing the task:

0. Locate yourself. — Find this SKILL.md's canonical path before editing.
1. What failed? — Fix the instruction that caused it.
2. What worked better than expected? — Promote to recommended practice.
3. What drifted? — Fix any script, reference, or dependency that no longer matches reality.
4. Log it. — Evolution-log entry with trigger, fix, and evidence.

Do NOT defer. The next invocation inherits whatever you leave behind.