This submission is for the ARC AGI challenge.
The "hilbert_predictor" incorporates several novel features and is under ongoing research.
If you're interested in collaborating or are working on ARC AGI and just want people to hang out with while you do it, we have a Discord here: https://discord.gg/SQy6vA5eE4
To train
python -m hilbert_predictor.train
To eval
python -m hilbert_predictor.eval
To train with wandb
wandb login
python -m hilbert_predictor.train --wandb
Tokens 0-9 are colors Token 10 is padding, ignored by the model Token 11 is START and token 12 is END token, which bookend the actual sequence