instruct based HuggingFace implementation for foundation attention-head.

Input

4381-dim embedding

Encoder

63 x Attention-UNet with 48 heads

Output

accuracy projection

gru based JAX implementation for token reward.

Input

2615-dim embedding

Encoder

13 x GAN with 46 heads

Output

mAP projection