use case · wild
Real text is messy. A single word can belong to several languages; a sentence can switch languages mid-way; a proper noun carries almost no signal. A good detector doesn't pretend — it spreads its probability, and that spread is the honest answer. Click a tricky example and read the distribution, not just the top label.
Notice how the top confidence drops and the runner-up bars rise as the input gets shorter or more mixed. For a product this is the cue to fall back — ask the user, wait for more text, or send it to a human — rather than trusting a coin-flip guess.
← Practical · Back to Language ID · Next: Multi-model →