I hosted my first eScriptorium instance

I hosted my first eScriptorium instance.

Setting it up self-hosted was nearly effortless. My earlier attempt at CVAT lasted a couple of weeks and the UI never felt intuitive.

I wired it to Kraken OCR and set up an iterative loop: the engine proposes a transcription, I fix the errors. This is much faster than annotating from scratch. Then I train the model on the corrected annotation and the loop repeats.

It really surprised me how fast the model adapted to Icelandic. After ~20 documents it was reliably reading the character set. It even picked up the Icelandic quotation marks „…", which SOTA LLMs still get wrong (try asking your favorite LLM about íslenskar gæsalappir).

Moving it off my laptop to a VPS was a Docker one-liner away. It now runs at escriptorium.sigurdurhaukur.com.

Next I hope to partner with the National Library of Iceland (Landsbókasafn Íslands – Háskólabókasafn) to crowd-source annotation over their archives and publish the first Icelandic OCR benchmark.

Anyone with experience running crowdsourced annotation projects? I'd love your advice on recruiting volunteers and keeping motivation up. Feel free to reach out via the contact page.