Free, legally downloadable PDFs that exercise different aspects of the pipeline.
| # | Book | Pages | Year | Best for testing |
|---|---|---|---|---|
| 1 | Understanding Deep Learning — Simon J.D. Prince | ~500 | 2024 | Math-heavy, figures, tables, deep TOC (parts/chapters/subsections) |
| 2 | Foundations of Large Language Models — Xiao et al. | ~150 | 2025 | Academic paper format, dense technical content, no embedded TOC |
| 3 | Software Design by Example (Python) — Greg Wilson | ~300 | 2024 | Code-heavy chapters, shell commands, structured TOC |
| 5 | Handbook of Software Engineering Methods — Oregon State | ~250 | 2024 | Structured textbook, tables, lists, multi-chapter |
| 6 | Mathematical Foundation of Reinforcement Learning — Shiyu Zhao | ~450 | 2024 | Heavy math (Bellman, policy gradient), per-chapter PDFs available, deep TOC |
| 15 | The Little Book of Deep Learning — François Fleuret | ~185 | 2023 | Rich 3-level TOC (parts/chapters/sections), front matter, compact |
Iconic papers — tests both embedded TOC and infer_toc font-analysis fallback.
| # | Paper | Pages | Year | TOC | Why it's iconic |
|---|---|---|---|---|---|
| 7 | Attention Is All You Need — Vaswani et al. | ~15 | 2017 | embedded | Introduced the Transformer architecture |
| 8 | BERT — Devlin et al. | ~16 | 2018 | — | Pre-train/fine-tune paradigm for NLP (masked LM) |
| 9 | Language Models are Few-Shot Learners (GPT-3) — Brown et al. | ~75 | 2020 | embedded | 175B params, few-shot without fine-tuning |
| 10 | Deep Residual Learning (ResNet) — He et al. | ~12 | 2015 | — | Skip connections, 100+ layer networks |
| 11 | An Image is Worth 16x16 Words (ViT) — Dosovitskiy et al. | ~22 | 2020 | — | Pure Transformer for image classification |
| 12 | NodeRAG — Xu et al. | ~20 | 2025 | — | Heterogeneous graph-based RAG framework |
| 13 | AlexNet — Krizhevsky et al. | ~9 | 2012 | none | Deep CNNs for ImageNet, launched modern deep learning |
| 14 | Adam Optimizer — Kingma & Ba | ~15 | 2014 | none | Default optimizer for most deep learning |
- Math & equations: #1 (Understanding Deep Learning), #6 (Math Foundation of RL)
- Code blocks & shell commands: #3 (Software Design by Example)
- Zero embedded TOC (pure
infer_toc): #13 (AlexNet), #14 (Adam) — tests font-analysis fallback end-to-end - Research papers with embedded TOC: #7 (Transformer), #9 (GPT-3) — tests embedded TOC on short/long papers
- Academic/arXiv format: #2 (Foundations of LLMs), #8, #10–#12 — various paper formats
- Rich embedded TOC (3 levels): #15 (Little Book of DL) — tests extract_toc, front matter classification, resolve_content_sections at level 1 and 2
- Structured textbook: #5 (Handbook of SE Methods) — tests cross-chapter coherence
- Large document handling: #1 (500+ pages) — tests auto-subdivision and context budget
- Short documents: #7, #8, #10 (~12–16 pages) — tests handling of compact single-section papers
| Book / Paper | License |
|---|---|
| Understanding Deep Learning | MIT Press, free PDF from author |
| Foundations of LLMs | arXiv open access |
| Software Design by Example | CC-BY-NC |
| Handbook of SE Methods | CC-BY-NC |
| Math Foundation of RL | Free PDF from author (Springer published version separate) |
| Research papers (#7–#12, #14) | arXiv open access (non-exclusive license to distribute) |
| AlexNet (#13) | NeurIPS proceedings, open access |
| Little Book of Deep Learning (#15) | Free PDF from author |