Nltk uses VADER method for sentiment analysis. And this method is rule based and does not have really good understanding of context.

You will get better result if you will try some Pre-Trained transformer based model such as Roberta. Because these model uses attention module and have really good understanding of context.

And you can also use free version of GPT or open source module such as LLama 3.1 8B to get the label.

I tried free version of GPT and I get accurate result on all 5 examples.