versae commited on
Commit
e8a9ff1
1 Parent(s): 51f880f

Prepping the run

Browse files
.gitattributes CHANGED
@@ -15,3 +15,5 @@
15
  *.pt filter=lfs diff=lfs merge=lfs -text
16
  *.pth filter=lfs diff=lfs merge=lfs -text
17
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
15
  *.pt filter=lfs diff=lfs merge=lfs -text
16
  *.pth filter=lfs diff=lfs merge=lfs -text
17
  *tfevents* filter=lfs diff=lfs merge=lfs -text
18
+ *.log filter=lfs diff=lfs merge=lfs -text
19
+ *.wandb filter=lfs diff=lfs merge=lfs -text
.gitignore ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ #run*.log
2
+ debug*.log
3
+ run*.wandb
4
+ wandb/
config.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "RobertaForMaskedLM"
4
+ ],
5
+ "attention_probs_dropout_prob": 0.1,
6
+ "bos_token_id": 0,
7
+ "eos_token_id": 2,
8
+ "gradient_checkpointing": false,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 768,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3072,
14
+ "layer_norm_eps": 1e-05,
15
+ "max_position_embeddings": 514,
16
+ "model_type": "roberta",
17
+ "num_attention_heads": 12,
18
+ "num_hidden_layers": 12,
19
+ "pad_token_id": 1,
20
+ "position_embedding_type": "absolute",
21
+ "transformers_version": "4.9.0.dev0",
22
+ "type_vocab_size": 1,
23
+ "use_cache": true,
24
+ "vocab_size": 50265
25
+ }
configs/base/config.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "RobertaForMaskedLM"
4
+ ],
5
+ "attention_probs_dropout_prob": 0.1,
6
+ "bos_token_id": 0,
7
+ "eos_token_id": 2,
8
+ "gradient_checkpointing": false,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 768,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3072,
14
+ "layer_norm_eps": 1e-05,
15
+ "max_position_embeddings": 514,
16
+ "model_type": "roberta",
17
+ "num_attention_heads": 12,
18
+ "num_hidden_layers": 12,
19
+ "pad_token_id": 1,
20
+ "position_embedding_type": "absolute",
21
+ "transformers_version": "4.9.0.dev0",
22
+ "type_vocab_size": 1,
23
+ "use_cache": true,
24
+ "vocab_size": 50265
25
+ }
configs/base/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
configs/large/config.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "RobertaForMaskedLM"
4
+ ],
5
+ "attention_probs_dropout_prob": 0.1,
6
+ "bos_token_id": 0,
7
+ "eos_token_id": 2,
8
+ "gradient_checkpointing": false,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 1024,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 4096,
14
+ "layer_norm_eps": 1e-05,
15
+ "max_position_embeddings": 514,
16
+ "model_type": "roberta",
17
+ "num_attention_heads": 16,
18
+ "num_hidden_layers": 24,
19
+ "pad_token_id": 1,
20
+ "position_embedding_type": "absolute",
21
+ "transformers_version": "4.9.0.dev0",
22
+ "type_vocab_size": 1,
23
+ "use_cache": true,
24
+ "vocab_size": 50265
25
+ }
configs/large/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
convert.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ import tempfile
3
+
4
+ import jax
5
+ from jax import numpy as jnp
6
+ from transformers import AutoTokenizer, FlaxRobertaForMaskedLM, RobertaForMaskedLM
7
+
8
+
9
+ def to_f32(t):
10
+ return jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype == jnp.bfloat16 else x, t)
11
+
12
+
13
+ def main():
14
+ # Saving extra files from config.json and tokenizer.json files
15
+ tokenizer = AutoTokenizer.from_pretrained("./")
16
+ tokenizer.save_pretrained("./")
17
+
18
+ # Temporary saving bfloat16 Flax model into float32
19
+ tmp = tempfile.mkdtemp()
20
+ flax_model = FlaxRobertaForMaskedLM.from_pretrained("./")
21
+ flax_model.params = to_f32(flax_model.params)
22
+ flax_model.save_pretrained(tmp)
23
+ # Converting float32 Flax to PyTorch
24
+ model = RobertaForMaskedLM.from_pretrained(tmp, from_flax=True)
25
+ model.save_pretrained("./", save_config=False)
26
+
27
+
28
+ if __name__ == "__main__":
29
+ main()
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
run_mlm_flax_stream.py ADDED
@@ -0,0 +1,780 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Team All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """
17
+ Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
18
+ text file or a dataset.
19
+
20
+ Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
21
+ https://huggingface.co/models?filter=masked-lm
22
+ """
23
+ import logging
24
+ import json
25
+ import os
26
+ import shutil
27
+ import sys
28
+ import time
29
+ from collections import defaultdict
30
+ from dataclasses import dataclass, field
31
+
32
+ # You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
33
+ import joblib
34
+ from pathlib import Path
35
+ from typing import Dict, List, Optional, Tuple
36
+
37
+ import datasets
38
+ import numpy as np
39
+ from datasets import load_dataset
40
+ from tqdm import tqdm
41
+
42
+ import flax
43
+ import jax
44
+ import jax.numpy as jnp
45
+ import kenlm # pip install https://github.com/kpu/kenlm/archive/master.zip
46
+ import optax
47
+ from flax import jax_utils, traverse_util
48
+ from flax.serialization import from_bytes, to_bytes
49
+ from flax.training import train_state
50
+ from flax.training.common_utils import get_metrics, onehot, shard
51
+ from transformers import (
52
+ CONFIG_MAPPING,
53
+ FLAX_MODEL_FOR_MASKED_LM_MAPPING,
54
+ AutoConfig,
55
+ AutoTokenizer,
56
+ FlaxAutoModelForMaskedLM,
57
+ HfArgumentParser,
58
+ PreTrainedTokenizerBase,
59
+ TensorType,
60
+ TrainingArguments,
61
+ is_tensorboard_available,
62
+ set_seed,
63
+ )
64
+
65
+
66
+ if datasets.__version__ <= "1.8.0":
67
+ raise ValueError("Make sure to upgrade `datasets` to a version >= 1.9.0 to use dataset streaming")
68
+
69
+
70
+ MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
71
+ MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
72
+
73
+
74
+ @dataclass
75
+ class ModelArguments:
76
+ """
77
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
78
+ """
79
+
80
+ model_name_or_path: Optional[str] = field(
81
+ default=None,
82
+ metadata={
83
+ "help": "The model checkpoint for weights initialization."
84
+ "Don't set if you want to train a model from scratch."
85
+ },
86
+ )
87
+ model_type: Optional[str] = field(
88
+ default=None,
89
+ metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
90
+ )
91
+ config_name: Optional[str] = field(
92
+ default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
93
+ )
94
+ tokenizer_name: Optional[str] = field(
95
+ default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
96
+ )
97
+ cache_dir: Optional[str] = field(
98
+ default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
99
+ )
100
+ use_fast_tokenizer: bool = field(
101
+ default=True,
102
+ metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
103
+ )
104
+ dtype: Optional[str] = field(
105
+ default="float32",
106
+ metadata={
107
+ "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
108
+ },
109
+ )
110
+
111
+ @dataclass
112
+ class DataTrainingArguments:
113
+ """
114
+ Arguments pertaining to what data we are going to input our model for training and eval.
115
+ """
116
+
117
+ dataset_name: Optional[str] = field(
118
+ default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
119
+ )
120
+ dataset_config_name: Optional[str] = field(
121
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
122
+ )
123
+ train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
124
+ validation_file: Optional[str] = field(
125
+ default=None,
126
+ metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
127
+ )
128
+ train_ref_file: Optional[str] = field(
129
+ default=None,
130
+ metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
131
+ )
132
+ validation_ref_file: Optional[str] = field(
133
+ default=None,
134
+ metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
135
+ )
136
+ overwrite_cache: bool = field(
137
+ default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
138
+ )
139
+ validation_split_percentage: Optional[int] = field(
140
+ default=5,
141
+ metadata={
142
+ "help": "The percentage of the train set used as validation set in case there's no validation split"
143
+ },
144
+ )
145
+ max_seq_length: Optional[int] = field(
146
+ default=None,
147
+ metadata={
148
+ "help": "The maximum total input sequence length after tokenization. Sequences longer "
149
+ "than this will be truncated. Default to the max input length of the model."
150
+ },
151
+ )
152
+ preprocessing_num_workers: Optional[int] = field(
153
+ default=None,
154
+ metadata={"help": "The number of processes to use for the preprocessing."},
155
+ )
156
+ mlm_probability: float = field(
157
+ default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
158
+ )
159
+ pad_to_max_length: bool = field(
160
+ default=False,
161
+ metadata={
162
+ "help": "Whether to pad all samples to `max_seq_length`. "
163
+ "If False, will pad the samples dynamically when batching to the maximum length in the batch."
164
+ },
165
+ )
166
+ line_by_line: bool = field(
167
+ default=False,
168
+ metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
169
+ )
170
+ text_column_name: str = field(
171
+ default="text", metadata={"help": "The name of the column to retrieve the training text."}
172
+ )
173
+ shuffle_buffer_size: int = field(
174
+ default=10000, metadata={"help": "The number of examples to pre-load for shuffling."}
175
+ )
176
+ num_train_steps: int = field(default=50000, metadata={"help": "The number of training steps."})
177
+ num_eval_samples: int = field(default=50000, metadata={"help": "The number of samples to be used for evaluation"})
178
+
179
+ def __post_init__(self):
180
+ if self.dataset_name is None and self.train_file is None and self.validation_file is None:
181
+ raise ValueError("Need either a dataset name or a training/validation file.")
182
+ else:
183
+ if self.train_file is not None:
184
+ extension = self.train_file.split(".")[-1]
185
+ assert extension in ["csv", "json", "jsonl", "txt", "gz"], "`train_file` should be a csv, a json (lines) or a txt file."
186
+ if self.validation_file is not None:
187
+ extension = self.validation_file.split(".")[-1]
188
+ assert extension in ["csv", "json", "jsonl", "txt", "gz"], "`validation_file` should be a csv, a json (lines) or a txt file."
189
+
190
+
191
+ @flax.struct.dataclass
192
+ class FlaxDataCollatorForLanguageModeling:
193
+ """
194
+ Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
195
+ are not all of the same length.
196
+
197
+ Args:
198
+ tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
199
+ The tokenizer used for encoding the data.
200
+ mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
201
+ The probability with which to (randomly) mask tokens in the input.
202
+
203
+ .. note::
204
+
205
+ For best performance, this data collator should be used with a dataset having items that are dictionaries or
206
+ BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
207
+ :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
208
+ argument :obj:`return_special_tokens_mask=True`.
209
+ """
210
+
211
+ tokenizer: PreTrainedTokenizerBase
212
+ mlm_probability: float = 0.15
213
+
214
+ def __post_init__(self):
215
+ if self.tokenizer.mask_token is None:
216
+ raise ValueError(
217
+ "This tokenizer does not have a mask token which is necessary for masked language modeling. "
218
+ "You should pass `mlm=False` to train on causal language modeling instead."
219
+ )
220
+
221
+ def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
222
+ # Handle dict or lists with proper padding and conversion to tensor.
223
+ batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
224
+
225
+ # If special token mask has been preprocessed, pop it from the dict.
226
+ special_tokens_mask = batch.pop("special_tokens_mask", None)
227
+
228
+ batch["input_ids"], batch["labels"] = self.mask_tokens(
229
+ batch["input_ids"], special_tokens_mask=special_tokens_mask
230
+ )
231
+ return batch
232
+
233
+ def mask_tokens(
234
+ self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
235
+ ) -> Tuple[jnp.ndarray, jnp.ndarray]:
236
+ """
237
+ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
238
+ """
239
+ labels = inputs.copy()
240
+ # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
241
+ probability_matrix = np.full(labels.shape, self.mlm_probability)
242
+ special_tokens_mask = special_tokens_mask.astype("bool")
243
+
244
+ probability_matrix[special_tokens_mask] = 0.0
245
+ masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
246
+ labels[~masked_indices] = -100 # We only compute loss on masked tokens
247
+
248
+ # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
249
+ indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
250
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
251
+
252
+ # 10% of the time, we replace masked input tokens with random word
253
+ indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
254
+ indices_random &= masked_indices & ~indices_replaced
255
+
256
+ random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
257
+ inputs[indices_random] = random_words[indices_random]
258
+
259
+ # The rest of the time (10% of the time) we keep the masked input tokens unchanged
260
+ return inputs, labels
261
+
262
+
263
+ @dataclass
264
+ class SamplingArguments:
265
+ """
266
+ Arguments pertaining to how to perform sampling of the dataset.
267
+ """
268
+
269
+ perplexity_model: Optional[str] = field(
270
+ default="./es.arpa.bin", metadata={"help": "Path to KenLM model to use to get perplexity values."}
271
+ )
272
+ sampling_method: Optional[str] = field(
273
+ default=None, metadata={"help": "Sample using a 'step' or 'gaussian' perplexity function per document, or 'random'."}
274
+ )
275
+ sampling_factor: Optional[float] = field(
276
+ default=None, metadata={"help": "Sampling factor. Integers for step function, decimals for gaussian."}
277
+ )
278
+ boundaries: Optional[str] = field(
279
+ default="536394.99320948,662247.50212365,919250.87225178", metadata={"help": "Quartile boundaries"}
280
+ )
281
+
282
+ def __post_init__(self):
283
+ self.boundaries = [float(q.strip()) for q in self.boundaries.split(",")]
284
+
285
+
286
+ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
287
+ num_samples = len(samples_idx)
288
+ samples_to_remove = num_samples % batch_size
289
+
290
+ if samples_to_remove != 0:
291
+ samples_idx = samples_idx[:-samples_to_remove]
292
+ sections_split = num_samples // batch_size
293
+ batch_idx = np.split(samples_idx, sections_split)
294
+ return batch_idx
295
+
296
+
297
+ def advance_iter_and_group_samples(train_iterator, num_samples, max_seq_length):
298
+ """
299
+ The training iterator is advanced so that after groupifying the samples,
300
+ `num_samples` of length `max_seq_length` are returned.
301
+ """
302
+ num_total_tokens = max_seq_length * num_samples
303
+ samples = defaultdict(list)
304
+
305
+ i = 0
306
+ while i < num_total_tokens:
307
+ tokenized_samples = next(train_iterator)
308
+ i += len(tokenized_samples["input_ids"])
309
+
310
+ # concatenate tokenized samples to list
311
+ samples = {k: samples[k] + tokenized_samples[k] for k in tokenized_samples.keys()}
312
+
313
+ # Concatenated tokens are split to lists of length `max_seq_length`.
314
+ # Note that remainedr of % max_seq_length are thrown away.
315
+ def group_texts(examples):
316
+ result = {
317
+ k: [t[i : i + max_seq_length] for i in range(0, num_total_tokens, max_seq_length)]
318
+ for k, t in examples.items()
319
+ }
320
+ return result
321
+
322
+ grouped_samples = group_texts(samples)
323
+ return grouped_samples
324
+
325
+
326
+ def write_train_metric(summary_writer, train_metrics, train_time, step):
327
+ summary_writer.scalar("train_time", train_time, step)
328
+
329
+ train_metrics = get_metrics(train_metrics)
330
+ for key, vals in train_metrics.items():
331
+ tag = f"train_{key}"
332
+ for i, val in enumerate(vals):
333
+ summary_writer.scalar(tag, val, step - len(vals) + i + 1)
334
+
335
+
336
+ def write_eval_metric(summary_writer, eval_metrics, step):
337
+ for metric_name, value in eval_metrics.items():
338
+ summary_writer.scalar(f"eval_{metric_name}", value, step)
339
+
340
+
341
+ def save_checkpoint_files(state, data_collator, training_args, save_dir):
342
+ unreplicated_state = jax_utils.unreplicate(state)
343
+ with open(os.path.join(save_dir, "optimizer_state.msgpack"), "wb") as f:
344
+ f.write(to_bytes(unreplicated_state.opt_state))
345
+ joblib.dump(training_args, os.path.join(save_dir, "training_args.joblib"))
346
+ joblib.dump(data_collator, os.path.join(save_dir, "data_collator.joblib"))
347
+ with open(os.path.join(save_dir, "training_state.json"), "w") as f:
348
+ json.dump({"step": unreplicated_state.step.item()}, f)
349
+
350
+
351
+ def restore_checkpoint(save_dir, state):
352
+ logger.info(f"Restoring checkpoint from {save_dir}")
353
+ with open(os.path.join(save_dir, "flax_model.msgpack"), "rb") as f:
354
+ params = from_bytes(state.params, f.read())
355
+
356
+ with open(os.path.join(save_dir, "optimizer_state.msgpack"), "rb") as f:
357
+ opt_state = from_bytes(state.opt_state, f.read())
358
+
359
+ args = joblib.load(os.path.join(save_dir, "training_args.joblib"))
360
+ data_collator = joblib.load(os.path.join(save_dir, "data_collator.joblib"))
361
+
362
+ with open(os.path.join(save_dir, "training_state.json"), "r") as f:
363
+ training_state = json.load(f)
364
+ step = training_state["step"]
365
+
366
+ return params, opt_state, step, args, data_collator
367
+
368
+
369
+ def rotate_checkpoints(path, max_checkpoints=5):
370
+ paths = sorted(Path(path).iterdir(), key=os.path.getmtime)[::-1]
371
+ if len(paths) > max_checkpoints:
372
+ for path_to_delete in paths[max_checkpoints:]:
373
+ try:
374
+ shutil.rmtree(path_to_delete)
375
+ except OSError:
376
+ os.remove(path_to_delete)
377
+
378
+
379
+ if __name__ == "__main__":
380
+ # See all possible arguments in src/transformers/training_args.py
381
+ # or by passing the --help flag to this script.
382
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
383
+
384
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, SamplingArguments))
385
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
386
+ # If we pass only one argument to the script and it's the path to a json file,
387
+ # let's parse it to get our arguments.
388
+ model_args, data_args, training_args, sampling_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
389
+ else:
390
+ model_args, data_args, training_args, sampling_args = parser.parse_args_into_dataclasses()
391
+
392
+ if (
393
+ os.path.exists(training_args.output_dir)
394
+ and os.listdir(training_args.output_dir)
395
+ and training_args.do_train
396
+ and not training_args.overwrite_output_dir
397
+ ):
398
+ raise ValueError(
399
+ f"Output directory ({training_args.output_dir}) already exists and is not empty."
400
+ "Use --overwrite_output_dir to overcome."
401
+ )
402
+
403
+ # Setup logging
404
+ logging.basicConfig(
405
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
406
+ level="INFO",
407
+ datefmt="[%X]",
408
+ )
409
+
410
+ # Log on each process the small summary:
411
+ logger = logging.getLogger(__name__)
412
+ logger.warning(
413
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
414
+ + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
415
+ )
416
+
417
+ # Set the verbosity to info of the Transformers logger (on main process only):
418
+ logger.info(f"Training/evaluation parameters {training_args}")
419
+
420
+ # Set seed before initializing model.
421
+ set_seed(training_args.seed)
422
+
423
+ # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
424
+ # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
425
+ # (the dataset will be downloaded automatically from the datasets Hub).
426
+ #
427
+ # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
428
+ # 'text' is found. You can easily tweak this behavior (see below).
429
+ if data_args.dataset_name is not None:
430
+ # Downloading and loading a dataset from the hub.
431
+ filepaths = {}
432
+ if data_args.train_file:
433
+ filepaths["train"] = data_args.train_file
434
+ if data_args.validation_file:
435
+ filepaths["validation"] = data_args.validation_file
436
+ try:
437
+ dataset = load_dataset(
438
+ data_args.dataset_name,
439
+ data_args.dataset_config_name,
440
+ cache_dir=model_args.cache_dir,
441
+ streaming=True,
442
+ split="train",
443
+ sampling_method=sampling_args.sampling_method,
444
+ sampling_factor=sampling_args.sampling_factor,
445
+ boundaries=sampling_args.boundaries,
446
+ perplexity_model=sampling_args.perplexity_model,
447
+ seed=training_args.seed,
448
+ data_files=filepaths,
449
+ )
450
+ except Exception as exc:
451
+ logger.warning(
452
+ f"Unable to load local dataset with perplexity sampling support. Using huggingface.co/datasets/{data_args.dataset_name}: {exc}"
453
+ )
454
+ dataset = load_dataset(
455
+ data_args.dataset_name,
456
+ data_args.dataset_config_name,
457
+ cache_dir=model_args.cache_dir,
458
+ streaming=True,
459
+ split="train",
460
+ )
461
+
462
+ if model_args.config_name:
463
+ config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
464
+ elif model_args.model_name_or_path:
465
+ config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
466
+ else:
467
+ config = CONFIG_MAPPING[model_args.model_type]()
468
+ logger.warning("You are instantiating a new config instance from scratch.")
469
+
470
+ if model_args.tokenizer_name:
471
+ tokenizer = AutoTokenizer.from_pretrained(
472
+ model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
473
+ )
474
+ elif model_args.model_name_or_path:
475
+ tokenizer = AutoTokenizer.from_pretrained(
476
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
477
+ )
478
+ else:
479
+ raise ValueError(
480
+ "You are instantiating a new tokenizer from scratch. This is not supported by this script."
481
+ "You can do it from another script, save it, and load it from here, using --tokenizer_name."
482
+ )
483
+
484
+ # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
485
+ # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
486
+ # efficient when it receives the `special_tokens_mask`.
487
+ def tokenize_function(examples):
488
+ return tokenizer(
489
+ examples[data_args.text_column_name],
490
+ return_special_tokens_mask=True
491
+ )
492
+
493
+ tokenized_datasets = dataset.map(
494
+ tokenize_function,
495
+ batched=True,
496
+ )
497
+
498
+ shuffle_seed = training_args.seed
499
+ tokenized_datasets = tokenized_datasets.shuffle(buffer_size=data_args.shuffle_buffer_size, seed=shuffle_seed)
500
+
501
+ # Enable tensorboard only on the master node
502
+ has_tensorboard = is_tensorboard_available()
503
+ if has_tensorboard and jax.process_index() == 0:
504
+ try:
505
+ # Enable Weight&Biases
506
+ import wandb
507
+ wandb.init(
508
+ entity='wandb',
509
+ project='hf-flax-bertin-roberta-es',
510
+ sync_tensorboard=True,
511
+ )
512
+ wandb.config.update(training_args)
513
+ wandb.config.update(model_args)
514
+ wandb.config.update(data_args)
515
+ from flax.metrics.tensorboard import SummaryWriter
516
+ summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
517
+ except ImportError as ie:
518
+ has_tensorboard = False
519
+ logger.warning(
520
+ f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
521
+ )
522
+ else:
523
+ logger.warning(
524
+ "Unable to display metrics through TensorBoard because the package is not installed: "
525
+ "Please run pip install tensorboard to enable."
526
+ )
527
+
528
+ # Data collator
529
+ # This one will take care of randomly masking the tokens.
530
+ data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
531
+
532
+ # Initialize our training
533
+ rng = jax.random.PRNGKey(training_args.seed)
534
+ dropout_rngs = jax.random.split(rng, jax.local_device_count())
535
+
536
+ if model_args.model_name_or_path:
537
+ model = FlaxAutoModelForMaskedLM.from_pretrained(
538
+ model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
539
+ )
540
+ else:
541
+ model = FlaxAutoModelForMaskedLM.from_config(
542
+ config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
543
+ )
544
+
545
+ # Store some constant
546
+ num_epochs = int(training_args.num_train_epochs)
547
+ train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
548
+ eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
549
+
550
+ # define number steps per stream epoch
551
+ num_train_steps = data_args.num_train_steps
552
+
553
+ # Create learning rate schedule
554
+ warmup_fn = optax.linear_schedule(
555
+ init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
556
+ )
557
+ decay_fn = optax.linear_schedule(
558
+ init_value=training_args.learning_rate,
559
+ end_value=0,
560
+ transition_steps=num_train_steps - training_args.warmup_steps,
561
+ )
562
+ linear_decay_lr_schedule_fn = optax.join_schedules(
563
+ schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
564
+ )
565
+
566
+ # We use Optax's "masking" functionality to not apply weight decay
567
+ # to bias and LayerNorm scale parameters. decay_mask_fn returns a
568
+ # mask boolean with the same structure as the parameters.
569
+ # The mask is True for parameters that should be decayed.
570
+ # Note that this mask is specifically adapted for FlaxBERT-like models.
571
+ # For other models, one should correct the layer norm parameter naming
572
+ # accordingly.
573
+ def decay_mask_fn(params):
574
+ flat_params = traverse_util.flatten_dict(params)
575
+ flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
576
+ return traverse_util.unflatten_dict(flat_mask)
577
+
578
+ # create adam optimizer
579
+ adamw = optax.adamw(
580
+ learning_rate=linear_decay_lr_schedule_fn,
581
+ b1=training_args.adam_beta1,
582
+ b2=training_args.adam_beta2,
583
+ eps=training_args.adam_epsilon,
584
+ weight_decay=training_args.weight_decay,
585
+ mask=decay_mask_fn,
586
+ )
587
+
588
+ # Setup train state
589
+ state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
590
+ saved_step = -1
591
+ if model_args.model_name_or_path and "checkpoint" in model_args.model_name_or_path:
592
+ params, opt_state, saved_step, args, data_collator = restore_checkpoint(model_args.model_name_or_path, state)
593
+ # Create learning rate schedule
594
+ warmup_fn = optax.linear_schedule(
595
+ init_value=0.0, end_value=args.learning_rate, transition_steps=args.warmup_steps
596
+ )
597
+ decay_fn = optax.linear_schedule(
598
+ init_value=args.learning_rate,
599
+ end_value=0,
600
+ transition_steps=data_args.num_train_steps - args.warmup_steps,
601
+ )
602
+ linear_decay_lr_schedule_fn = optax.join_schedules(
603
+ schedules=[warmup_fn, decay_fn], boundaries=[args.warmup_steps]
604
+ )
605
+ # create adam optimizer
606
+ adamw = optax.adamw(
607
+ learning_rate=linear_decay_lr_schedule_fn,
608
+ b1=training_args.adam_beta1,
609
+ b2=training_args.adam_beta2,
610
+ eps=training_args.adam_epsilon,
611
+ weight_decay=args.weight_decay,
612
+ mask=decay_mask_fn,
613
+ )
614
+ state = train_state.TrainState(
615
+ step=saved_step,
616
+ apply_fn=model.__call__,
617
+ params=params,
618
+ tx=adamw,
619
+ opt_state=opt_state,
620
+ )
621
+ # self.args = args
622
+ # data_collator = data_collator
623
+ # scheduler_fn = args.learning_rate
624
+ model.params = params
625
+
626
+
627
+ # Define gradient update step fn
628
+ def train_step(state, batch, dropout_rng):
629
+ dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
630
+
631
+ def loss_fn(params):
632
+ labels = batch.pop("labels")
633
+
634
+ logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
635
+
636
+ # compute loss, ignore padded input tokens
637
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
638
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
639
+
640
+ # take average
641
+ loss = loss.sum() / label_mask.sum()
642
+
643
+ return loss
644
+
645
+ grad_fn = jax.value_and_grad(loss_fn)
646
+ loss, grad = grad_fn(state.params)
647
+ grad = jax.lax.pmean(grad, "batch")
648
+ new_state = state.apply_gradients(grads=grad)
649
+
650
+ metrics = jax.lax.pmean(
651
+ {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
652
+ )
653
+
654
+ return new_state, metrics, new_dropout_rng
655
+
656
+ # Create parallel version of the train step
657
+ p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
658
+
659
+ # Define eval fn
660
+ def eval_step(params, batch):
661
+ labels = batch.pop("labels")
662
+
663
+ logits = model(**batch, params=params, train=False)[0]
664
+
665
+ # compute loss, ignore padded input tokens
666
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
667
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
668
+
669
+ # compute accuracy
670
+ accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
671
+
672
+ # summarize metrics
673
+ metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
674
+ metrics = jax.lax.psum(metrics, axis_name="batch")
675
+
676
+ return metrics
677
+
678
+ p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
679
+
680
+ # Replicate the train state on each device
681
+ state = jax_utils.replicate(state)
682
+
683
+ train_time = 0
684
+ train_start = time.time()
685
+ train_metrics = []
686
+ eval_metrics = []
687
+
688
+ training_iter = iter(tokenized_datasets)
689
+
690
+ max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
691
+ eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
692
+
693
+ steps = tqdm(range(num_train_steps), desc="Training...", position=0)
694
+ for step in range(0, num_train_steps):
695
+ if step < saved_step:
696
+ steps.update(1)
697
+ continue
698
+ # ======================== Training ================================
699
+ try:
700
+ samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
701
+ except StopIteration:
702
+ # Once the end of the dataset stream is reached, the training iterator
703
+ # is reinitialized and reshuffled and a new eval dataset is randomely chosen.
704
+ shuffle_seed += 1
705
+ tokenized_datasets.set_epoch(shuffle_seed)
706
+
707
+ training_iter = iter(tokenized_datasets)
708
+
709
+ eval_dataset = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
710
+ samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
711
+
712
+ # process input samples
713
+ model_inputs = data_collator(samples, pad_to_multiple_of=16)
714
+
715
+ # Model forward
716
+ model_inputs = shard(model_inputs.data)
717
+ state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
718
+
719
+ train_metrics.append(train_metric)
720
+
721
+ if step % training_args.logging_steps == 0 and step > 0:
722
+ steps.write(
723
+ f"Step... ({step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
724
+ )
725
+ train_time += time.time() - train_start
726
+ if has_tensorboard and jax.process_index() == 0:
727
+ write_train_metric(summary_writer, train_metrics, train_time, step)
728
+ train_metrics = []
729
+
730
+ # ======================== Evaluating ==============================
731
+ if step % training_args.eval_steps == 0 and step > 0:
732
+ eval_samples_idx = jnp.arange(data_args.num_eval_samples)
733
+ eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
734
+
735
+ for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=1)):
736
+ # process input samples
737
+ batch_eval_samples = {k: [v[idx] for idx in batch_idx] for k, v in eval_samples.items()}
738
+ model_inputs = data_collator(batch_eval_samples, pad_to_multiple_of=16)
739
+
740
+ # Model forward
741
+ model_inputs = shard(model_inputs.data)
742
+ metrics = p_eval_step(state.params, model_inputs)
743
+ eval_metrics.append(metrics)
744
+
745
+ # normalize eval metrics
746
+ eval_metrics = get_metrics(eval_metrics)
747
+ eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
748
+ eval_normalizer = eval_metrics.pop("normalizer")
749
+ eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
750
+
751
+ # Update progress bar
752
+ steps.desc = f"Step... ({step + 1}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
753
+
754
+ if has_tensorboard and jax.process_index() == 0:
755
+ write_eval_metric(summary_writer, eval_metrics, step)
756
+ eval_metrics = []
757
+
758
+ # save checkpoint after eval_steps
759
+ if step % training_args.save_steps == 0 and step > 0 and jax.process_index() == 0:
760
+ logger.info(f"Saving checkpoint at {step} steps")
761
+ params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
762
+ model.save_pretrained(
763
+ training_args.output_dir,
764
+ params=params,
765
+ push_to_hub=training_args.push_to_hub,
766
+ commit_message=f"Saving weights and logs of step {step + 1}",
767
+ )
768
+ save_checkpoint_files(state, data_collator, training_args, training_args.output_dir)
769
+ checkpoints_dir = Path(training_args.output_dir) / "checkpoints" / f"checkpoint-{step}"
770
+ checkpoints_dir.mkdir(parents=True, exist_ok=True)
771
+ model.save_pretrained(checkpoints_dir, params=params)
772
+ save_checkpoint_files(state, data_collator, training_args, checkpoints_dir)
773
+ rotate_checkpoints(
774
+ Path(training_args.output_dir) / "checkpoints",
775
+ max_checkpoints=training_args.save_total_limit
776
+ )
777
+
778
+ # update tqdm bar
779
+ steps.update(1)
780
+
run_stream.128.sh ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # From https://arxiv.org/pdf/1907.11692.pdf for base model
2
+ python -c "import jax; print('TPUs', jax.device_count())"
3
+ python ./run_mlm_flax_stream.py \
4
+ --output_dir="./outputs" \
5
+ --model_type="roberta" \
6
+ --config_name="./configs/base" \
7
+ --tokenizer_name="./configs/base" \
8
+ --dataset_name="./mc4" \
9
+ --dataset_config_name="es" \
10
+ --train_file="../mc4-es-train-50M-random.jsonl" \
11
+ --max_seq_length="128" \
12
+ --pad_to_max_length \
13
+ --per_device_train_batch_size="256" \
14
+ --per_device_eval_batch_size="256" \
15
+ --adam_beta1="0.9" \
16
+ --adam_beta2="0.98" \
17
+ --adam_epsilon="1e-6" \
18
+ --learning_rate="6e-4" \
19
+ --weight_decay="0.01" \
20
+ --save_steps="10000" \
21
+ --save_total_limit="5" \
22
+ --warmup_steps="24000" \
23
+ --overwrite_output_dir \
24
+ --num_train_steps="250000" \
25
+ --eval_steps="10000" \
26
+ --dtype="bfloat16" \
27
+ --logging_steps="500" 2>&1 | tee run_stream.log
run_stream.512.sh ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # From https://arxiv.org/pdf/1907.11692.pdf for base model
2
+ python -c "import jax; print('TPUs', jax.device_count())"
3
+ python ./run_mlm_flax_stream.py \
4
+ --model_name_or_path="bertin-project/bertin-base-gaussian" \
5
+ --output_dir="./outputs" \
6
+ --model_type="roberta" \
7
+ --config_name="./configs/base" \
8
+ --tokenizer_name="./configs/base" \
9
+ --dataset_name="versae/mc4-es-sampled" \
10
+ --dataset_config_name="gaussian" \
11
+ --max_seq_length="512" \
12
+ --pad_to_max_length \
13
+ --per_device_train_batch_size="48" \
14
+ --per_device_eval_batch_size="48" \
15
+ --adam_beta1="0.9" \
16
+ --adam_beta2="0.98" \
17
+ --adam_epsilon="1e-6" \
18
+ --learning_rate="6e-4" \
19
+ --weight_decay="0.01" \
20
+ --save_steps="1000" \
21
+ --save_total_limit="5" \
22
+ --warmup_steps="500" \
23
+ --overwrite_output_dir \
24
+ --num_train_steps="50000" \
25
+ --eval_steps="1000" \
26
+ --dtype="bfloat16" \
27
+ --logging_steps="500" 2>&1 | tee run_stream.log
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>", "add_prefix_space": false, "errors": "replace", "sep_token": "</s>", "cls_token": "<s>", "pad_token": "<pad>", "mask_token": "<mask>", "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "RobertaTokenizer"}
vocab.json ADDED
The diff for this file is too large to render. See raw diff