Three synthetic datasets covering different types of bias grouped by target, namely sexism, racism and ableism. The reason for distinguishing the records by abuse targets is due to the need for specialised datasets addressing different phenomena of abusive language with a fine-grained approach. The resulting data do not contain samples from datasets under license: the contents we release are therefore freely available. Briefly, the first dataset on sexism contains 1,200 non-hateful and 4,423 hateful samples; the second one on racism contains 400 non-hateful and 1,500 hateful records; the last one on ableism contains 220 hateful sentences. The label distribution is radically different from traditional abusive language datasets, where the prevalent class is non-hateful. This choice is motivated by the fact that we want to mainly focus on the phenomena surrounding social prejudices providing realistic and diverse examples, with the aim of exploring in depth the language used to convey biases.