← Back to News

Dataset · 12 Nov 2026

An open test for filling in missing signs

Kanichira, Singam, Katyal and Somala (IIT Hyderabad) · Zenodo 24164070

Illustration: a clay sealing with one sign broken away, the gap outlined and a candidate sign shown in it.
Editorial illustration, not a real object or a transcription.

In a sentence

A team released code, data and predictions for a model that guesses damaged or missing signs.

In plain words

Many seals are chipped, so some signs are lost. This team trained a model to fill the gaps and published everything: the code, how they split the texts for training and testing, and the model’s guesses. Because it is all open, anyone can check how good it really is.

Technical detail

Neural masked-sign restoration with public train/test splits and predictions (CC BY 4.0 and MIT, about 6 MB). The record page does not describe the corpus or the model. The main risk is leakage: near-copies of the same seal text can sit on both sides of a split, which inflates scores.

What we take from it

Tested

We graded it. 36% of its test texts have a near-copy in its training texts. On clean texts the neural model does no better than a simple three-sign model (28.5% vs 29.8% right on the first guess).

The idea we are borrowing

Grade the new sign-restoration model fairly, with near-copies kept out of its test set

Where it stands: Tested.

Read our result in the repository.

← Back to News