Sentences sampled from news articles around the year 2020 (forget exactly which year). 1400 annotated sentences, with chosen named entities covered by [POS], [NEG] or [NEU] masks, based on how the sentence portrays them. The task is Aspect Based Sentiment Recognition, in the domain of political coverage.
Five of my friends volunteered sentences from their news reading to send me these using a browser plugin annotating tool I made.
Some annotators are better than others. None the less, this small amount of data allowed a successful fine tuning for an ABSA BERT model, with which, semi supervised learning showed pretty impressive results.
There is almost 0 quality control, so things may get messy.