A few misc notes:<p>1. The better way to get all Hacker News data instead of blasting the API is to download the data from the official BigQuery dataset, which can do the task in a single query: <a href="https://news.ycombinator.com/item?id=40644563">https://news.ycombinator.com/item?id=40644563</a><p>2. For labeling the posts, instead of label-then-explanation, it may be better to do explanation-then-label to give the model a chance to reason though the edge cases.<p>3. Following up from #2, for prompt engineering the system prompt, it would likely be better to give a list of multiple valid examples and invalid examples (as noted after the fact) to guide reasoning.<p>4. Since the target label is a binary objective, it may be more practical/faster/cheaper to create a normal logistic regression model (e.g. tf-idf/BoW) from a large representative sample, then use that to predict the rest of the labels.<p>The more advanced way to do #4 would be to encode the posts as text embeddings first then use them as the input for a small MLP model...which I may or may not have a project in the pipeline based around that approach.