Dataset for text classification

I am going to collect tweets for sentiment analysis (pos,neg,neu) to train a model based on BERT to find out the polarity of the tweet over a specific topic. As there is not any dataset with any label in this regard, based on some hashtags, we divide tweets into positive and negative classes, however we do not have idea about what kind of tweets can be in dataset for the Neutral s. Is there any advise that how we can have training text for Neutral classes?