Flan-mini

Given the enormous size of the Flan Collection, we opted to work with a carefully selected subset that maintains a high level of task diversity while reducing the overall dataset size. In the Table below, we present the specific tasks included in our subset of Flan, along with their respective dataset sizes. As the public release of the Flan Collection does not include programming tasks, we augment the collection with existing code datasets. Specifically, we include CodeContests, APPS, and CodeSearchNet. Following the data processing pipeline of Flan Collection, we sample a fixed number of examples from each dataset, where each example is randomly augmented with different prompt templates. Specifically, the examples are processed with a pool of handcrafted prompt templates and may be used as zero-shot examples or grouped together with few-shot demonstrations. We incorporated various ChatGPT datasets, including Alpaca, Code Alpaca, and ShareGPT, into our Flan-mini collection.