| import os |
| import json |
| from transformers import PreTrainedTokenizer |
|
|
|
|
| class TwentyQTokenizer(PreTrainedTokenizer): |
| """Byte-level tokenizer for TwentyQ. |
| |
| Also stores the question and target vocabularies — these are the model's |
| "tokens" in the same way that a language model's tokenizer stores its vocabulary. |
| """ |
|
|
| vocab_files_names = {"vocab_file": "vocab.json"} |
|
|
| def __init__(self, vocab_file=None, **kwargs): |
| self.questions = [] |
| self.targets = [] |
| if vocab_file and os.path.exists(vocab_file): |
| with open(vocab_file) as f: |
| data = json.load(f) |
| self.questions = data.get("questions", []) |
| self.targets = data.get("targets", []) |
|
|
| self._byte_vocab = {i: chr(i) if 32 <= i < 127 else f"<0x{i:02X}>" for i in range(256)} |
| self._byte_vocab[256] = "<pad>" |
| self._byte_vocab[257] = "<s>" |
| self._byte_vocab[258] = "</s>" |
| self._str_to_id = {v: k for k, v in self._byte_vocab.items()} |
|
|
| kwargs.setdefault("pad_token", "<pad>") |
| kwargs.setdefault("bos_token", "<s>") |
| kwargs.setdefault("eos_token", "</s>") |
| kwargs.setdefault("model_max_length", 4096) |
| super().__init__(vocab_file=vocab_file, **kwargs) |
|
|
| @property |
| def vocab_size(self): |
| return 259 |
|
|
| def get_vocab(self): |
| return dict(self._str_to_id) |
|
|
| def _tokenize(self, text): |
| return [self._byte_vocab.get(b, f"<0x{b:02X}>") for b in text.encode("utf-8")] |
|
|
| def _convert_token_to_id(self, token): |
| return self._str_to_id.get(token, 0) |
|
|
| def _convert_id_to_token(self, index): |
| return self._byte_vocab.get(index, "<0x00>") |
|
|
| def convert_tokens_to_string(self, tokens): |
| byte_vals = [] |
| for t in tokens: |
| if t in ("<pad>", "<s>", "</s>"): |
| continue |
| if t.startswith("<0x") and t.endswith(">"): |
| byte_vals.append(int(t[3:-1], 16)) |
| elif len(t) == 1: |
| byte_vals.append(ord(t)) |
| return bytes(byte_vals).decode("utf-8", errors="replace") |
|
|
| def save_vocabulary(self, save_directory, filename_prefix=None): |
| vocab_file = os.path.join( |
| save_directory, |
| (filename_prefix + "-" if filename_prefix else "") + "vocab.json", |
| ) |
| with open(vocab_file, "w") as f: |
| json.dump({"questions": self.questions, "targets": self.targets}, f, indent=2) |
| return (vocab_file,) |
|
|
