metadata
license: apache-2.0
Don't use this model for any applied task. It too small to be practically useful. It is just a part of a weird research project.
An extremely small version of T5 with these parameters
"d_ff": 1024,
"d_kv": 64,
"d_model": 256,
"num_heads": 4,
"num_layers": 1, # yes, just one layer
The model was pre-trained on realnewslike subset of C4 for 1 epoch with sequence length 64. Corresponding WandB run: click.