Blog
Lector pays the data back
Lector now pays $100 a month, or 10% of profit, whichever is greater, to the open data projects its language packs depend on. Working out who to pay turned up a surprise: about half the list has no way to receive money at all — including Anki, and including the one dependency that has stopped being maintained entirely.
Mandarin, Japanese and Korean: three scripts, three different problems
Three East Asian packs are live. I expected one hard job and got three unrelated ones: Mandarin needed the words found, Japanese needed a morphological analyser after the segmenter shredded every verb, and Korean needed its grammar peeled off the end of each word. What each fix bought, measured.
Three new language packs: Ukrainian, Polish and Czech
Ukrainian shipped because a learner pointed out that reading Ukrainian under the Russian pack mistranslates words and cuts others in half. Polish and Czech followed. What each pack measures, and where the gaps still are.
Import YouTube transcripts and podcasts
Two new ways to bring native audio and video into Lector: a YouTube video's transcript with clickable timestamps, or a podcast you read along to as it plays — both with full word-state tracking and vocabulary mining.
Which local LLM translates best? A reproducible eval
A reproducible benchmark of 24 on-device, self-hosted, and cloud models translating into English — with the low-resource case (Afrikaans) front and centre. The headline: a local 18 GB model ties frontier cloud on Afrikaans→English.
What I learned by OCRing an Afrikaans dictionary
Digitising a 346-page print dictionary to widen Lector's offline vocabulary — why a smaller book beat a bigger wordlist, what OCR does to a language built on diacritics, and what the data says about Afrikaans.
Run translations locally with LM Studio
Point Lector's translation agent at a local LLM through LM Studio's OpenAI-compatible API — no API keys, no cloud. Setup, the Docker networking gotcha, and headless config.