BBF on Atari 100k, for TensorFlow.js
Snapshots of BBF ("Bigger, Better, Faster", Schwarzer et al. 2023) trained on Atari 100k with the released BBF.gin
(replay ratio 64, 2 updates per move), one seed per game, saved every 10,000 moves. Used by the page
Two Hours of Atari.
<game>/<step>/policy.json describes the tensors in policy.bin (target network; conv kernels float16,
dense kernels int8 with one scale per output column). test.json holds reference Q values from JAX on real game states.