jsjol.github.io

Personal research website for Jens Sjölund, Assistant Professor of Machine Learning at Uppsala University. Live at https://jsjol.github.io.

Built with Jekyll on a heavily reduced fork of the al-folio theme. Pushing to master triggers the Deploy site workflow, which builds the site, purges unused CSS and publishes _site to the gh-pages branch.

Local development

Requires Ruby with Bundler, and ImageMagick for the responsive image variants:

brew install imagemagick   # provides `magick`
bundle install
bundle exec jekyll serve    # http://localhost:4000

The build needs nothing else — no Python, no Node. A clean build takes about five seconds and should print no warnings; treat any warning as a defect.

Before committing, format the sources:

npm install                 # first time only
npx prettier . --write

Where the content lives

What	Where
About page	_pages/about.md
News items	_news/
Publications	_bibliography/papers.bib
Research threads	_projects/
Courses	_teachings/
Group members	_data/people.yml
Social links	_data/socials.yml
Repository cards	_data/repositories.yml
CV	assets/pdf/cv.pdf, linked from _pages/cv.md
Citation counts	_data/citations.yml, regenerated by CI (see below)
Scholar exclusions	_data/scholar_ignore.yml

Adding a news item

Create _news/YYYY-MM-DD-slug.md whose only front matter is its date:

---
date: 2026-08-18
---

New paper at [Venue]: [**Title**](link). [One sentence.] Work led by [Author].

News items are rendered inline in the news table on / and /news/, so they get no standalone pages. Keep the filename date and the front-matter date equal.

Adding a publication

Append a BibTeX entry to _bibliography/papers.bib. Beyond the standard fields, the templates understand abbr, abstract, arxiv, award, award_name, bibtex_show, code, html, pdf, poster, preview, selected, slides, supp, video and website. selected={true} puts the entry on the front page; preview={file.png} looks for that file in assets/img/publication_preview/.

Adding an image

Nothing needs to be wider than 1400 px — that is the largest variant _includes/figure.liquid puts in the srcset. Compress before committing:

magick in.png -strip -resize '1400x1400>' -quality 85 out.png
pngquant --quality=70-92 --strip --force --output out.png out.png

Workflows

• Deploy site — on every push to master that touches site content.
• Update publications — Mon/Wed/Fri, in two jobs of deliberately different trust levels. See below.

Keeping the publication list current

bin/sync_publications.py diffs the Google Scholar profile against _bibliography/papers.bib. Run it with no arguments for a report:

python bin/sync_publications.py                    # report only
python bin/sync_publications.py --annotate         # attach Scholar ids
python bin/sync_publications.py --append           # add missing papers

Existing entries are never rewritten. The bibliography is hand-curated — preview in particular — so the script may only insert a google_scholar_id line into an entry that lacks one, and append new entries at the end of the file. Both are asserted afterwards: --append checks that the previous content is a byte prefix of the result, and --annotate checks that the file grew by exactly the inserted lines. No field value is ever touched.

The Update publications workflow splits this by how much review each half needs. The refresh job updates citation counts and attaches Scholar ids, and commits straight to master because both are mechanically safe. The propose job appends papers that are on Scholar but not in the bibliography and opens a pull request, because a new entry needs its venue and author list checked and wants an abbr, a preview and possibly selected added by hand.

The citation counts need a local run

bin/update_scholar_citations.py takes about ten seconds here:

python bin/update_scholar_citations.py    # needs `scholarly`, see requirements.txt

It reads the profile page directly — one request, standard library only, about three seconds — and falls back to scholarly only if that fails. The old scholarly-only path issued several requests per run, and Google answers datacenter addresses with a CAPTCHA, which is why the scheduled workflow produced nothing for seven weeks while reporting success. The direct path stands a better chance from CI but is not guaranteed either, so the workflow warns on each failure and turns red once _data/citations.yml is more than ten days old. When that happens, run the command above and push.

The two paths were checked against each other: both return 107 records and 3444 citations, with no difference in any id, count or title.

Two things worth knowing:

• Detection costs no Scholar requests. bin/update_scholar_citations.py already downloads the whole publication list with a single author-level scholarly.fill(), so the sync only reads _data/citations.yml. It never calls scholarly.fill(pub) per paper, which is what triggers Scholar's CAPTCHA and stalls fetch-publications partway through the list. Metadata for a new paper comes from Crossref, and from the arXiv API when Crossref has no record — which is also why the generated BibTeX keys carry a real first-author surname rather than unknown.
• Scholar sometimes attributes someone else's paper to the profile. Add its title or publication id to _data/scholar_ignore.yml and it will not be proposed again. Supervised PhD theses are listed there too: they belong on the group page via _data/people.yml, not in the publication list.
• OpenAlex is a second detection source, and is report-only. It works from CI even when Scholar does not, but it keeps separate records for the preprint and the published version of the same paper, leaves HTML in titles, and lists theses, so its findings are printed for a human to check rather than appended. Titles are matched exactly, then on a prefix (for Scholar's truncated titles), then on word overlap above 0.72 — which is what recognises "A linear programming approach to inverse planning in radiosurgery" as the entry already in the file under "… in Gamma Knife radiosurgery".

Where several Scholar records map onto one entry — the patent families all share titles like "Methods for inverse planning" — the script attaches no id and says so, rather than guessing which record belongs to which entry.

Local changes to the theme

This fork has diverged from upstream al-folio deliberately, so updates have to be cherry-picked rather than merged. The substantive differences:

• The blog, bookshelf, CV-from-JSON, distill, repository-trophy, comment, newsletter and search features are removed, along with their layouts, includes, Sass partials, data files, assets and CDN libraries.
• MDBootstrap is gone. It cost 277 kB of CSS and 288 kB of JS to supply four shadow rules, which now live in _sass/_utilities.scss.
• _plugins/cache-bust.rb replaces the jekyll-cache-bust gem, whose CSS digest hashed a directory that does not exist in this repository, so the ?v= query string never changed and browsers served stale CSS.
• An empty title: in _config.yml names the site after its owner. Upstream requires the literal string blank for this and renders an empty <h1> and an empty navbar brand otherwise.
• _pages/people.md uses a CSS grid defined in _sass/_people.scss rather than Bootstrap row/column utilities.