One coder put out a full data set this week. The project goes by CCORE GLB-3D-Coder. It trains small AI models to make 3D shapes.
The coder built a small model first. It has 4 billion settings. It writes code that draws 3D scenes. Each scene saves as a GLB file. Other 3D tools can open that file right away.
This week's drop adds the data used to train that model. The set holds 3,918 cases across 40 kinds of items. Those items span people, cars, lab tools, and musical gear.
Each case was drawn and checked before it went out. Each saved file was checked too. The set splits into 3,501 cases for training and 417 for tests. Big labs rarely give out this much of their own recipe.
The data set carries an open MIT license. Anyone can grab it and train their own model. The small model built from it sees real use now. It has 12,206 downloads and 191 likes on Hugging Face.
The Reddit post about the drop got seven notes. Coders asked if the set would grow to new kinds of items. The coder wrote back to each one in the thread.
One coder used the model in an earlier post to build a small web game. That game used shapes the model made on its own: people, props, and ground. It shows what a tiny model can do with the right data behind it.
Big labs tend to guard their training data closely. This drop goes the other way. It hands over the full recipe along with the model itself.
That kind of openness has a cost. Other coders can now copy the exact steps behind the model. This coder chose to make that trade anyway. Open recipes like this help the whole small-model crowd move faster together.