Just spitballing, but maybe a good tradeoff is to use NLP to find good candidate comments that are likely to contain a sentiment, and then analysing a small number of them with a more expensive model (say a quantized 3B or 7B LLM with a good prompt). The quality over quantity approach