Help with 0 accuracy on training a textcat

[info2000]

I'm training a textcat model, with 3 categories and these example's amount:
cat1: train 184, eval 33
cat2: 4545, 248
cat3: 2379, 110

and these cat_micro scores:
cat1: f1: 0, precision 0, recall 0
cat2 0.87
cat3 0.83

the topic for cat1 is clear different than for cat2 and cat3, so overlap examples is rare
I increased model width up to 512, but that only improved cat2 and cat3
Training from es_core_news_lg and only textcat pipeline
eval frecuency 200
patience 4000
batch size 2000 words

what could be the next parameter to touch? I can't have many more examples for cat1

Thanks