A char-RNN is a bad way to go about it. The results are effectively random - there's a ton of structure in recipe-space, but forcing it all through the bottleneck of text (and then sampling!) is going to hide a lot of it. You want something like "Sukiyaki in French style: A novel system for transformation of dietary patterns" https://arxiv.org/abs/1705.03487, Kazama et al 2017's `word2vec` (using structured recipes like https://open.blogs.nytimes.com/2015/04/09/extracting-structu... ); or IBM Watson's recipe generation ( https://www.eater.com/2014/2/26/6272773/food-truck-is-actual... https://www.newyorker.com/magazine/2016/11/28/cooking-with-c... ).