minbpe-hs is a port of Andrej Karpathy's concise byte-level byte pair encoding (BPE) implementation, minbpe, to Haskell. During training, BPE compresses a sequence by repeatedly finding the most frequent pair of elements in the input and merging them into a new token whilst maintaining a record of the merges and the tokens. Encoding is performed by recursively merging pairs with the smallest merge index until there remain no pairs to be merged, and encoded representations can be decoded backed into the original through vocabulary look-ups. Given its recursive nature, BPE lends itself naturally to functional programming, and the objective of this repository is to offer an alternative functional formulation of it that is orthogonal to typical implementations in imperative languages such as Python or Rust to aid learners in gaining a firmer grasp on BPE by viewing it through a different lens.
Similar to minbpe, BPE.Basic and BPE.Regex both supply the same fundamental functionalities of training, encoding, and decoding, but the former treats a text corpus as one continuous sequence whereas the latter, in addition to supporting special tokens, divides it into smaller segments using a regex pattern and considers merges only within these segments. BPE.Base contains core operations employed by the two tokenizers and also provides storing & loading utilities. There is no counterpart to minbpe's GPT4Tokenizer as tiktoken does not expose a Haskell frontend.
Questions and feedback are welcome in the comments sections.
