Sanjin2024's picture
Update README.md
df833aa verified
metadata
license: apache-2.0

TinyRecursiveModels - Maze Hard (Attention-based)

This model is a reproduction of the TinyRecursiveModels project by Samsung SAIL Montreal, specifically trained and evaluated on the Maze Hard (30×30) reasoning task.

image

image

8 × H200, global_batch_size=4608, Runtime: 2h

Test Accuracy: 78.70%