conv_kernel_size: 7
d_embed: 128
d_model: 256
d_output: 256
dropout: 0.0
ffn_mult: 2.6666666666666665
matryoshka_dims:
- 32
- 64
- 128
- 256
max_seq_len: 1024
n_heads: 4
n_layers: 6
n_random_features: 128
pooling: mean
spatial_rank: 32
variant: transformer
vocab_size: 30000