This tokenizer was trained on a small corpus of concatenated ARPAbet pronunciation tokens + punctuation from the python g2p_en library computed over the entire synthbot/pony-speech dataset and 240k lines from generics_kb_best, from community-datasets/generics_kb.
i.e. But one on one, let's clean it. -> BAH1T WAH1N AA1N WAH1N , LEH1TS KLIY1N IH1T . Uses BPE with vocab size of 1024.
It is trained on the same data as https://huggingface.co/therealvul/tokenizer_g2pen with the following differences:
- It does not split on whitespace as a token
- It uses ByteLevel in pretokenization/decoding step
- It uses a token vocab size of 1024 instead of 384