/paperify

Transform any document, web page, or eBook into a research paper (ChatGPT not required)

Primary LanguageShellMIT LicenseMIT

Paperify

Paperify transforms any document, web page, or ebook into a research paper.

The text of the generated paper is the same as the text of the original document, but figures and equations from real papers are interspersed throughout.

A paper title and abstract are added (optionally generated by ChatGPT, if you provide an API key), and the entire paper is compiled with the IEEE $\LaTeX$ template for added realism.

example

Install

First, install the dependencies (or use Docker):

  • curl
  • Python 3
  • Pandoc
  • jq
  • LaTeX (via TeXLive)
  • ImageMagick (optional)

For example, on Debian-based systems (e.g., Debian, Ubuntu, Kali, WSL):

sudo apt update
sudo apt install --no-install-recommends \
  pandoc \
  curl ca-certificates \
  jq \
  python3 \
  imagemagick \
  texlive texlive-publishers texlive-science lmodern texlive-latex-extra

Then, clone the repo (or directly pull the script), and execute it.

curl -L https://github.com/jstrieb/paperify/raw/master/paperify.sh \
  | sudo tee /usr/local/bin/paperify
sudo chmod +x /usr/local/bin/paperify

paperify -h

Examples

  • examples/cox.pdf

    Convert Russ Cox's transcript of Doug McIlroy's talk on the history of Bell Labs into a paper saved to the /tmp/ directory as article.pdf.

    paperify \
  --from-format html \
  "https://research.swtch.com/bell-labs" \
  /tmp/article.pdf

  • examples/london.pdf

    Download figures and equations from the 1000 latest computer science papers on arXiv.org. Intersperse the figures and equations into Jack London's Call of the Wild with a higher-than-default equation frequency. Use ChatGPT to generate a paper title, author, abstract, and metadata for an imaginary paper on soft body robotics. Save the file in the current directory as london.pdf.

    paperify \
  --arxiv-category cs \
  --num-papers 1000 \
  --equation-frequency 18 \
  --chatgpt-token "sk-[REDACTED]" \
  --chatgpt-topic "soft body robotics" \
  "https://standardebooks.org/ebooks/jack-london/the-call-of-the-wild/downloads/jack-london_the-call-of-the-wild.epub" \
  london.pdf

Docker

Alternatively, run Paperify from within a Docker container. To run the first example from within Docker and build to ./build/cox.pdf:

docker run \
  --rm \
  -it \
  --volume "$(pwd)/build":/root/build \
  jstrieb/paperify \
    --from-format html \
    "https://research.swtch.com/bell-labs" \
    build/cox.pdf

Usage

usage: paperify [OPTIONS] <URL or path> <output file>

OPTIONS:
  --temp-dir <DIR>            Directory for assets (default: /tmp/paperify)
  --from-format <FORMAT>      Format of input file (default: input suffix)
  --arxiv-category <CAT>      arXiv.org paper category (default: math)
  --num-papers <NUM>          Number of papers to download (default: 100)
  --max-parallelism <PROCS>   Maximum simultaneous processes (default: 32)
  --figure-frequency <N>      Chance of a figure is 1/N per paragraph (default: 25)
  --equation-frequency <N>    Chance of an equation is 1/N per paragraph (default: 25)
  --max-size <BYTES>          Max allowed image size in bytes (default 2500000)
  --min-equation-length <N>   Minimum equation length in characters (default 5)
  --max-equation-length <N>   Maximum equation length in characters (default 120)
  --min-caption-length <N>    Minimum figure caption length in characters (default 20)
  --chatgpt-token <TOKEN>     ChatGPT token to generate paper title, abstract, etc.
  --chatgpt-topic <TOPIC>     Paper topic ChatGPT will generate metadta for
  --quiet                     Don't log statuses
  --skip-downloading          Don't download papers from arXiv.org
  --skip-extracting           Don't extract equations and captions
  --skip-metadata             Don't regenerate metadata
  --skip-filtering            Don't filter out large files or non-diagram images

Note that the --skip-* flags are useful when you have already run the script once and do not want to repeat the process of downloading and extracting data.

Known Issues

  • Images with query parameters in the src URL of some web pages are extracted by Pandoc with the query parameters in the filename, and LaTeX gives errors about "unknown file extension" when compiling.
  • Papers may contain images that are not diagrams, such as portraits of the authors or institution logos. Paperify uses a highly imperfect heuristic to remove these if the convert command line tool is present: only images with white, nearly-white, or transparent pixels in the top left and bottom right corners are kept. This works surprisingly well, but there are always some false positives and false negatives.
  • Non-ASCII Unicode characters cannot be processed by pdflatex, and will be stripped before the PDF is compiled.
  • Paperify uses Markdown as a (purposefully) lossy intermediate representation for documents before they are converted to LaTeX. As a result, information and styling from the original may be stripped.
  • A handful of papers contain huge numbers of images. The ones that do this also tend to have some of the worst images. Images can be manually pruned from the /tmp/paperify/images directory, and the same command can be re-run with the --skip-* flags to rebuild the paper using new figures and equations.
  • Different systems install different LaTeX packages. If you're missing packages, you may want to bite the bullet and apt install texlive-full. It's very big, but it's got everything you'll ever need in there.
  • Figure captions usually have nothing to do with figures themselves.
  • No matter how convincing a paper may appear, anyone looking over your shoulder who actually reads the words will know very quickly that something is off.
  • Side effects of reading the code include nausea, dizziness, confusion, bleeding from the eyes, and deep love/hatred for the creators of Unix pipelines.

How to Read the Code

In general, I'm a proponent of reading (or at least skimming) code before you run it, when possible. Usually, my code is written to be read. In this case, not so much.

Apologies in advance to anyone who tries to read the code. It started as four very cursed lines of Bash (without line wrapping) that I attempted to clean up a little. It is now many more than four lines of Bash, most of which remain very cursed. The small Python portion is particularly hard on the eyes, though it may possess a grotesque beauty for true functional programmers.

Everything is in paperify.sh. It can be read top-to-bottom or bottom-to-top, and there is a fat LaTeX template as a heredoc smack in the middle.

Project Status

Strange as it may sound, this project is complete. I want to live in a world where working software doesn't always grow until it becomes a Lovecraftian spaghetti monster.

I have added every feature that I wanted to add. It does what I wanted it to do, as well as I wanted it to do it. No further development required.

As such, I will try to address issues opened on GitHub, but I do not expect to address feature requests. I may merge pull requests.

Even if there are no recent commits, I'm hopeful that this script will continue to work many years from now.

Greetz & Acknowledgments

Greetz to several unnamed friends who offered helpful commentary prior to release.

Special shout out to the friends who suggested, as a follow-up project, making a browser extension to transform the current web page into a scientific paper. Sort of like Firefox reader mode, but for viewing Twitter when someone looking over your shoulder expects you to be doing something else.

Thanks to arXiv.org for hosting tons of papers with LaTeX source to mine.

Greetz to Project Gutenberg, Standard Ebooks, and Alexandra Elbakyan.

Lovingly released on Labor Day 2023; dedicated to procrastinating laborers of knowledge.