File size: 13,288 Bytes
f3a6aa4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 | """HuggingFace Transformers-compatible wrapper for JSONTokenizer.
Provides JSONPreTrainedTokenizer, a PreTrainedTokenizer subclass that
wraps JSONTokenizer for use with the HuggingFace ecosystem:
- save_pretrained / from_pretrained
- AutoTokenizer.from_pretrained (with trust_remote_code=True)
- tokenizer(json_string) -> BatchEncoding
- Padding, truncation, batch processing, return_tensors
Requires: pip install json-tokenizer[huggingface]
"""
from __future__ import annotations
import json
import os
from typing import Any, Dict, List, Optional, Tuple, Union
try:
from transformers import PreTrainedTokenizer
except ImportError:
raise ImportError(
"The HuggingFace transformers library is required for this module. "
"Install it with: pip install json-tokenizer[huggingface]"
)
from json_tokenizer.tokenizer import JSONTokenizer, StructuralTokens
from json_tokenizer.bpe import BPETrainer
VOCAB_FILES_NAMES = {"vocab_file": "json_tokenizer_vocab.json"}
# Structural token ID -> HF-compatible string name.
# Uses <name> format which cannot collide with BPE tokens because
# the BPE pre-tokenizer splits <, >, : into separate tokens.
_STRUCTURAL_TOKEN_NAMES = {
StructuralTokens.PAD: "<pad>",
StructuralTokens.START: "<s>",
StructuralTokens.END: "</s>",
StructuralTokens.OBJ_START: "<obj_start>",
StructuralTokens.OBJ_END: "<obj_end>",
StructuralTokens.ARR_START: "<arr_start>",
StructuralTokens.ARR_END: "<arr_end>",
StructuralTokens.COLON: "<colon>",
StructuralTokens.COMMA: "<comma>",
StructuralTokens.NULL: "<null>",
StructuralTokens.TRUE: "<true>",
StructuralTokens.FALSE: "<false>",
StructuralTokens.STR_DELIM: "<str_delim>",
StructuralTokens.NUM_PREFIX: "<num_prefix>",
StructuralTokens.KEY_PREFIX: "<key_prefix>",
StructuralTokens.UNK: "<unk>",
}
_STRUCTURAL_NAME_TO_ID = {v: k for k, v in _STRUCTURAL_TOKEN_NAMES.items()}
class JSONPreTrainedTokenizer(PreTrainedTokenizer):
"""HuggingFace-compatible wrapper around JSONTokenizer.
Usage:
# From a trained JSONTokenizer:
tok = JSONTokenizer(bpe_vocab_size=4096)
tok.train(data)
hf_tok = JSONPreTrainedTokenizer.from_json_tokenizer(tok)
# Encode/decode via HF API:
output = hf_tok('{"name": "Alice", "age": 30}')
print(output["input_ids"])
print(hf_tok.decode(output["input_ids"]))
# Save and reload:
hf_tok.save_pretrained("./my_tokenizer")
loaded = JSONPreTrainedTokenizer.from_pretrained("./my_tokenizer")
"""
vocab_files_names = VOCAB_FILES_NAMES
model_input_names = ["input_ids", "attention_mask"]
def __init__(
self,
vocab_file: Optional[str] = None,
unk_token: str = "<unk>",
bos_token: str = "<s>",
eos_token: str = "</s>",
pad_token: str = "<pad>",
**kwargs,
):
# Internal state β populated from vocab_file or from_json_tokenizer
if not hasattr(self, "_json_tokenizer"):
self._json_tokenizer: Optional[JSONTokenizer] = None
if not hasattr(self, "_hf_vocab"):
self._hf_vocab: Dict[str, int] = {}
if not hasattr(self, "_hf_id_to_token"):
self._hf_id_to_token: Dict[int, str] = {}
if vocab_file is not None and os.path.isfile(vocab_file):
self._load_vocab_file(vocab_file)
super().__init__(
unk_token=unk_token,
bos_token=bos_token,
eos_token=eos_token,
pad_token=pad_token,
**kwargs,
)
# ββ Factory ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@classmethod
def from_json_tokenizer(
cls, tokenizer: JSONTokenizer, **kwargs
) -> "JSONPreTrainedTokenizer":
"""Create from a trained JSONTokenizer instance.
Args:
tokenizer: A trained JSONTokenizer.
**kwargs: Additional arguments passed to __init__.
Returns:
A new JSONPreTrainedTokenizer wrapping the provided tokenizer.
"""
if not tokenizer._trained:
raise ValueError("JSONTokenizer must be trained before wrapping.")
instance = cls.__new__(cls)
instance._json_tokenizer = tokenizer
instance._hf_vocab = {}
instance._hf_id_to_token = {}
instance._build_hf_vocab()
instance.__init__(vocab_file=None, **kwargs)
return instance
# ββ Vocab building βββββββββββββββββββββββββββββββββββββββββββββββββ
def _load_vocab_file(self, vocab_file: str) -> None:
"""Reconstruct a JSONTokenizer from our saved vocab file."""
with open(vocab_file, "r", encoding="utf-8") as f:
data = json.load(f)
config = data["config"]
tok = JSONTokenizer(
bpe_vocab_size=config["bpe_vocab_size"],
max_key_vocab=config["max_key_vocab"],
min_key_freq=config["min_key_freq"],
bpe_min_freq=config["bpe_min_freq"],
)
tok._key_to_id = {k: int(v) for k, v in data["key_vocab"].items()}
tok._id_to_key = {int(v): k for k, v in data["key_vocab"].items()}
tok._key_offset = config["key_offset"]
tok._bpe_offset = config["bpe_offset"]
bpe_data = data["bpe_model"]
bpe = BPETrainer(
vocab_size=bpe_data["vocab_size"],
min_frequency=bpe_data["min_frequency"],
)
bpe.merges = [tuple(m) for m in bpe_data["merges"]]
bpe.vocab = bpe_data["vocab"]
bpe._id_to_tok = None
tok._bpe = bpe
tok._build_vocab_lookup()
tok._trained = True
self._json_tokenizer = tok
self._build_hf_vocab()
def _build_hf_vocab(self) -> None:
"""Build the unified {token_string: id} mapping across all tiers."""
tok = self._json_tokenizer
self._hf_vocab = {}
self._hf_id_to_token = {}
# Structural tokens (0-15)
for tid, name in _STRUCTURAL_TOKEN_NAMES.items():
self._hf_vocab[name] = tid
self._hf_id_to_token[tid] = name
# Reserved tokens (16-31)
for tid in range(16, StructuralTokens.RESERVED_END):
name = f"<reserved_{tid}>"
self._hf_vocab[name] = tid
self._hf_id_to_token[tid] = name
# Key vocabulary tokens
for key_str, tid in tok._key_to_id.items():
name = f"<key:{key_str}>"
self._hf_vocab[name] = tid
self._hf_id_to_token[tid] = name
# BPE tokens
for bpe_token, bpe_local_id in tok._bpe.vocab.items():
full_id = tok._bpe_offset + bpe_local_id
# Collision guard (only <UNK> from BPE could theoretically collide)
if bpe_token in self._hf_vocab:
bpe_token_name = f"bpe:{bpe_token}"
else:
bpe_token_name = bpe_token
self._hf_vocab[bpe_token_name] = full_id
self._hf_id_to_token[full_id] = bpe_token_name
# ββ Required PreTrainedTokenizer overrides βββββββββββββββββββββββββ
@property
def vocab_size(self) -> int:
if self._json_tokenizer is None:
return len(_STRUCTURAL_TOKEN_NAMES)
return self._json_tokenizer.vocab_size
def get_vocab(self) -> Dict[str, int]:
vocab = dict(self._hf_vocab)
vocab.update(self.added_tokens_encoder)
return vocab
def _tokenize(self, text: str, **kwargs) -> List[str]:
"""Tokenize a JSON string into HF token strings.
The HF pipeline calls: tokenize(text) -> _tokenize -> list[str]
then convert_tokens_to_ids maps those to IDs.
We parse the JSON, encode via JSONTokenizer (skipping START/END
since HF adds special tokens via build_inputs_with_special_tokens),
then convert IDs to our HF token string names.
"""
if self._json_tokenizer is None:
return [self.unk_token]
try:
ids = self._json_tokenizer.encode(text)
except (ValueError, json.JSONDecodeError):
# Not valid JSON β encode as raw string via BPE
ids = [StructuralTokens.START]
self._json_tokenizer._encode_string(text, ids)
ids.append(StructuralTokens.END)
# Strip START/END β HF adds them via build_inputs_with_special_tokens
if ids and ids[0] == StructuralTokens.START:
ids = ids[1:]
if ids and ids[-1] == StructuralTokens.END:
ids = ids[:-1]
return [self._hf_id_to_token.get(tid, self.unk_token) for tid in ids]
def _convert_token_to_id(self, token: str) -> int:
return self._hf_vocab.get(
token, self._hf_vocab.get(self.unk_token, StructuralTokens.UNK)
)
def _convert_id_to_token(self, index: int) -> str:
return self._hf_id_to_token.get(index, self.unk_token)
def convert_tokens_to_string(self, tokens: List[str]) -> str:
"""Reconstruct a JSON string from token strings.
Converts token strings -> IDs, wraps with START/END,
and delegates to JSONTokenizer.decode().
"""
if self._json_tokenizer is None:
return ""
ids = [StructuralTokens.START]
for token in tokens:
tid = self._convert_token_to_id(token)
ids.append(tid)
ids.append(StructuralTokens.END)
try:
return self._json_tokenizer.decode(ids)
except Exception:
return " ".join(tokens)
# ββ Special tokens βββββββββββββββββββββββββββββββββββββββββββββββββ
def build_inputs_with_special_tokens(
self,
token_ids_0: List[int],
token_ids_1: Optional[List[int]] = None,
) -> List[int]:
"""Wrap with START (bos) and END (eos) tokens."""
bos = [self.bos_token_id]
eos = [self.eos_token_id]
if token_ids_1 is None:
return bos + token_ids_0 + eos
return bos + token_ids_0 + eos + bos + token_ids_1 + eos
def get_special_tokens_mask(
self,
token_ids_0: List[int],
token_ids_1: Optional[List[int]] = None,
already_has_special_tokens: bool = False,
) -> List[int]:
"""1 for special tokens (START/END), 0 for content tokens."""
if already_has_special_tokens:
return super().get_special_tokens_mask(
token_ids_0=token_ids_0,
token_ids_1=token_ids_1,
already_has_special_tokens=True,
)
if token_ids_1 is None:
return [1] + [0] * len(token_ids_0) + [1]
return (
[1] + [0] * len(token_ids_0) + [1]
+ [1] + [0] * len(token_ids_1) + [1]
)
def create_token_type_ids_from_sequences(
self,
token_ids_0: List[int],
token_ids_1: Optional[List[int]] = None,
) -> List[int]:
"""Segment IDs: 0 for first sequence, 1 for second."""
bos_eos = 2 # one bos + one eos
if token_ids_1 is None:
return [0] * (len(token_ids_0) + bos_eos)
return [0] * (len(token_ids_0) + bos_eos) + [1] * (len(token_ids_1) + bos_eos)
# ββ Persistence ββββββββββββββββββββββββββββββββββββββββββββββββββββ
def save_vocabulary(
self,
save_directory: str,
filename_prefix: Optional[str] = None,
) -> Tuple[str]:
"""Save the vocabulary to a single JSON file.
This file contains everything needed to reconstruct the
JSONTokenizer: config, key vocab, and BPE model.
"""
if not os.path.isdir(save_directory):
raise ValueError(f"Not a directory: {save_directory}")
vocab_file = os.path.join(
save_directory,
(filename_prefix + "-" if filename_prefix else "")
+ VOCAB_FILES_NAMES["vocab_file"],
)
tok = self._json_tokenizer
data = {
"version": "json-tokenizer-hf-v1",
"config": {
"bpe_vocab_size": tok.bpe_vocab_size,
"max_key_vocab": tok.max_key_vocab,
"min_key_freq": tok.min_key_freq,
"bpe_min_freq": tok.bpe_min_freq,
"key_offset": tok._key_offset,
"bpe_offset": tok._bpe_offset,
},
"key_vocab": tok._key_to_id,
"bpe_model": {
"vocab_size": tok._bpe.vocab_size,
"min_frequency": tok._bpe.min_frequency,
"merges": [list(m) for m in tok._bpe.merges],
"vocab": tok._bpe.vocab,
},
}
with open(vocab_file, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2, ensure_ascii=False)
return (vocab_file,)
|