---
title: Tessera - span-level language ID
emoji: 🧩
colorFrom: indigo
colorTo: green
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: Languages a document-level LID would have discarded
---

# Tessera

Document-level language identification gives **one language per page**. Real pages are
not one language, and the languages that get thrown away are the ones that cannot
spare the loss.

Paste a page. The demo labels every span and shows what a single document-level label
would have cost.

Everything runs **in your browser**: a 17 MB float16 export of the model, with the
featuriser and the bidirectional GRU ported to JavaScript and verified against the
Python reference to 3e-04. Nothing is uploaded anywhere.

For the three built-in examples, the document-level verdicts shown are **measured** by
running GlotLID v3, OpenLID and fastText lid.176 on those exact strings offline. On
the first example GlotLID returns `anp_Deva` (Angika) at p=0.974 for a page containing
no Devanagari at all, and fastText discards the page entirely.

**Tessera runs behind GlotLID / OpenLID, not instead of them.**

- Code: https://github.com/NagaYu/tessera
- Model: https://huggingface.co/NagaYu/tessera-span-lid
- Dataset: https://huggingface.co/datasets/NagaYu/tessera-mixed-pages