Pushing open data from inside a legal publisher (2019): two pro bono partnerships in France & Luxembourg
Announcements: Etalab announcement - Luxembourg press release 3
TL;DR. From spring to autumn 2019, we ran two pro bono partnerships to help open up court decisions. With Etalab (the French government’s open data unit within DINUM, the interministerial digital directorate) and the Cour de cassation (France’s supreme court for civil and criminal matters) we shared methodology and, above all, made Flair fast enough for large-scale anonymization (about 10× faster: ~30 days -> <3 days on a single 2080Ti for instance). With the Luxembourg Prosecutor General’s Office (the national public prosecutor’s office), we tested the whole pipeline end-to-end, produced a working PoC, and handed it over to de-risk future rollout. I’m genuinely happy we did both - pushing open data from an editor isn’t obvious, but it’s the direction the field is taking and the results were concrete. (Etalab, Justice Luxembourg)
UPDATE 2022: You can see the follow-through: by July 4, 2022, the Prosecutor General launched JUANO, an application to help pseudonymize decisions for publication. Different team, different time - but a clear continuation of the same arc. Communiqué de presse
Why this matters (and why do it from an editor?)
I’ve said it elsewhere: I believe in the virtues of open data. It attracts talent, speeds up research, and lets real-world users build on top. From a publisher’s seat, that vision isn’t automatic - you have to align people on the fact that opening up is positive-sum for everyone, especially at a time when some “modern” legaltech try to sell what should be open. Our bet was simple: contribute engineering and methods in the open, and let the ecosystem benefit.
Partnership #1 - Etalab × Lefebvre Sarrut (Cour de cassation), spring 2019
We signed with Etalab (the French government’s open data taskforce within DINUM) to collaborate together, which led to publish our pseudonymization code under a free license and to work with the Cour de cassation (France’s supreme court) on the methodology of pseudonymization. The point was to compare methods and accelerate what worked. Practically: Etalab bridged our team and the Court’s team; approaches could be combined if that improved results.
What we brought was largely engineering. On our side we kept the model intact and rewired how computation was done in Flair. Net effect: ~10× faster inference (1m16s -> 11s per 100 cases) which turns a full inventory pass from ~30 days into <3 days on a single 2080Ti - the kind of small box you actually have. That changes the operational game: re-processing the whole backlog becomes a week-end task, not a month-long project. This is why engineering sits at the heart of applied ML.
Style-wise, this aligned well with the Court’s strong, dedicated team on the subject (via Etalab’s OpenJustice/EIG effort): we shared methodology, made the speed-ups available, and let their team push accuracy and evaluation on their side. Positive-sum, in public.
Partnership #2 - Luxembourg Prosecutor General’s Office, autumn 2019
In November we signed a six-month, no-money partnership with the Parquet général (Luxembourg’s Prosecutor General’s Office - the national public prosecutor’s office; with Larcier Luxembourg): evaluate AI anonymization for mass publication and accelerate their open-justice policy.
Here we ran the pipeline end-to-end - there wasn’t a standing anonymization team yet - so we focused on getting something that worked in their context, measured it, and then handed the PoC back to the institution. The goal was explicit: de-risk the “can this run here?” question so the office could staff and scale later. (Martine Solovieff was Prosecutor General at the time.)
How we convinced people internally
- Align on the “why.” This is where the world is going: open decisions, modern tooling, reproducible pipelines. The job was to align everyone on the fact that it’s a good thing with positive spillovers - for institutions, for citizens, and yes, for an editor operating in a healthier ecosystem.
- Reassure by measurement. Publish code, publish metrics (speed, accuracy, throughput), and show the delta in plain numbers. Reliability is not a promise; it’s a benchmark and a CI run.
What actually came out (beyond slides)
- Open code & methods that others could evaluate and build on (Etalab/Cour de cassation).
- A 10× speed-up that made “full-inventory anonymization in a couple of days” realistic on commodity GPUs.
- An in-context PoC at the Luxembourg Prosecutor General’s Office that answered the hard, local questions before any procurement or staffing. (And later, a homegrown tool.)
Pro bono always carries the usual risk: no money -> no value. Here, both projects produced real effects - usable code, measurable performance, and institutional learning that stuck. That was the point.
If you want the technical backstory on anonymization itself, the two long posts from that period cover it: why we switched from spaCy to Flair and made it 10× faster, and a benchmark (spaCy, Flair, mBERT, CamemBERT) on messy, OCR’d commercial cases.