I went to the lodash docs and asked about how I'd use the 'pipeline' operator (which doesn't exist) and it correctly pointed out that pipeline isn't a thing, and suggested chain() for normal code and flow() for lodash fp instead. That's pretty much spot on. If I was guessing I'd suggest that the base model has a lot more lodash code examples in the training data, which probably makes a big difference to the quality of the output.