Besides using the “shuffle” buffer, we also shuffle the shards (underlying data files; dataset.n_shards returns the number of them) for more randomness, which leads to a different result unless the seed is fixed.
1 Like