Albanisht was created to make communication easier between generations and Albanians of the diaspora, and for Albanian and foreign travellers who want to understand one another when they need to, when they'd like to, or simply to learn something.

Today's AI models are good enough that what's missing is the cultural and linguistic layer built specifically for Albanian. That's what's here: an open test of the translation, failed cases included; Bisedo, where two people take turns on one phone; and a dataset that grows from corrections sent by Albanian speakers.

This is a non-profit family project. All it aims to be is the tool that's been missing.

Methodology

A frontier language model combined with retrieval over a curated parallel corpus and a hand-built linguistic knowledge base. For every translation request we find the most similar examples in the corpus, look up the idioms and false friends that apply to the text, and build a prompt tuned specifically for Albanian.

Twelve rules for reading Albanian cover the admirative mood, aspect, non-active voice, kinship, Gheg forms, blessings and idioms, religious register, false friends and word order. Fourteen norms of the Albanian Academy's orthography and grammar govern written Albanian, followed by a self-check for agreement, case and calques. A parallel corpus of 1,117 Albanian–English sentence pairs from Tatoeba plus 177 hand-curated pairs reviewed by a native speaker. A retrieval index keyed by character n-grams. A dialect detector that tells Tosk from Gheg.

Sources we cite: the Academy of Sciences of Albania; Newmark, Hubbard and Prifti, Standard Albanian. A Reference Grammar; Camaj, Albanian Grammar; Mëniku and Campos, Discovering Albanian; Friedman's work on the Balkan sprachbund; the Tatoeba sentence corpus; Albanian Wiktionary.

Each directive for a feature is mapped to its source grammar; each corpus entry is tagged by origin (akademia / newmark / camaj / sister-review / user-report).

Known limitations

An AI judge grades the benchmark. A Claude model reads each translation against reference answers and scores meaning; a fixed rule in our code turns that into pass or fail, the same for every system. The judge is a Claude model grading Claude output, so self-preference is possible, and its agreement with native speakers is still being measured. Every output and every reason for a fail is on the benchmark page.
Google Translate answers are captured by hand. Google updates its engine continuously. Between our capture dates, its answers may differ from what we show.
Benchmark size. 98 hand-written test cases catch regressions on the features we care most about, and 160 sentences from Meta's FLORES-200 benchmark, which no prompt was tuned on, check longer real-world text. Neither is exhaustive. We extend them as real use shows us new failure modes.
Test answers are static. The lists of acceptable answers were written at a point in time. Language is alive. We revisit them every year.

Get in touch

General questions: support@albanisht.com. Security reports: security@albanisht.com.