gchhablani commited on
Commit
439ed65
1 Parent(s): 8712bf8

Update README

Browse files
Files changed (1) hide show
  1. README.md +20 -4
README.md CHANGED
@@ -9,6 +9,7 @@ datasets:
9
  - wikipedia
10
  ---
11
  # MultiBERTs Seed 11 (uncased)
 
12
  Seed 11 MultiBERTs (pretrained BERT) model on English language using a masked language modeling (MLM) objective. It was introduced in
13
  [this paper](https://arxiv.org/pdf/2106.16163.pdf) and first released in
14
  [this repository](https://github.com/google-research/language/tree/master/language/multiberts). This model is uncased: it does not make a difference
@@ -17,10 +18,12 @@ between english and English.
17
  Disclaimer: The team releasing MultiBERTs did not write a model card for this model so this model card has been written by [gchhablani](https://hf.co/gchhablani).
18
 
19
  ## Model description
 
20
  MultiBERTs models are transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it
21
  was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of
22
  publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
23
  was pretrained with two objectives:
 
24
  - Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run
25
  the entire masked sentence through the model and has to predict the masked words. This is different from traditional
26
  recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like
@@ -29,11 +32,12 @@ was pretrained with two objectives:
29
  - Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes
30
  they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to
31
  predict if the two sentences were following each other or not.
32
- This way, the model learns an inner representation of the English language that can then be used to extract features
33
- useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard
34
- classifier using the features produced by the MultiBERTs model as inputs.
35
 
36
  ## Intended uses & limitations
 
37
  You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
38
  be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=multiberts) to look for
39
  fine-tuned versions on a task that interests you.
@@ -42,10 +46,12 @@ to make decisions, such as sequence classification, token classification or ques
42
  generation you should look at model like GPT2.
43
 
44
  ### How to use
 
45
  Here is how to use this model to get the features of a given text in PyTorch:
 
46
  ```python
47
  from transformers import BertTokenizer, BertModel
48
- tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
49
  model = BertModel.from_pretrained("multiberts-seed-11")
50
  text = "Replace me by any text you'd like."
51
  encoded_input = tokenizer(text, return_tensors='pt')
@@ -53,39 +59,48 @@ output = model(**encoded_input)
53
  ```
54
 
55
  ### Limitations and bias
 
56
  Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
57
  predictions. This bias will also affect all fine-tuned versions of this model. For an understanding of bias of this particular
58
  checkpoint, please try out this checkpoint with the snippet present in the [Limitation and bias section](https://huggingface.co/bert-base-uncased#limitations-and-bias) of the [bert-base-uncased](https://huggingface.co/bert-base-uncased) checkpoint.
59
 
60
  ## Training data
 
61
  The MultiBERTs models were pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038
62
  unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and
63
  headers).
 
64
  ## Training procedure
65
 
66
  ### Preprocessing
 
67
  The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
68
  then of the form:
 
69
  ```
70
  [CLS] Sentence A [SEP] Sentence B [SEP]
71
  ```
 
72
  With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in
73
  the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a
74
  consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two
75
  "sentences" has a combined length of less than 512 tokens.
76
  The details of the masking procedure for each sentence are the following:
 
77
  - 15% of the tokens are masked.
78
  - In 80% of the cases, the masked tokens are replaced by `[MASK]`.
79
  - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
80
  - In the 10% remaining cases, the masked tokens are left as is.
81
 
82
  ### Pretraining
 
83
  The model was trained on 16 Cloud TPU v2 chips for two million steps with a batch size
84
  of 256. The sequence length was set to 512 throughout. The optimizer
85
  used is Adam with a learning rate of 1e-4, \\(\beta_{1} = 0.9\\) and \\(\beta_{2} = 0.999\\), a weight decay of 0.01,
86
  learning rate warmup for 10,000 steps and linear decay of the learning rate after.
87
 
88
  ### BibTeX entry and citation info
 
89
  ```bibtex
90
  @article{DBLP:journals/corr/abs-2106-16163,
91
  author = {Thibault Sellam and
@@ -112,6 +127,7 @@ learning rate warmup for 10,000 steps and linear decay of the learning rate afte
112
  bibsource = {dblp computer science bibliography, https://dblp.org}
113
  }
114
  ```
 
115
  <a href="https://huggingface.co/exbert/?model=multiberts">
116
  <img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
117
  </a>
 
9
  - wikipedia
10
  ---
11
  # MultiBERTs Seed 11 (uncased)
12
+
13
  Seed 11 MultiBERTs (pretrained BERT) model on English language using a masked language modeling (MLM) objective. It was introduced in
14
  [this paper](https://arxiv.org/pdf/2106.16163.pdf) and first released in
15
  [this repository](https://github.com/google-research/language/tree/master/language/multiberts). This model is uncased: it does not make a difference
 
18
  Disclaimer: The team releasing MultiBERTs did not write a model card for this model so this model card has been written by [gchhablani](https://hf.co/gchhablani).
19
 
20
  ## Model description
21
+
22
  MultiBERTs models are transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it
23
  was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of
24
  publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
25
  was pretrained with two objectives:
26
+
27
  - Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run
28
  the entire masked sentence through the model and has to predict the masked words. This is different from traditional
29
  recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like
 
32
  - Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes
33
  they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to
34
  predict if the two sentences were following each other or not.
35
+ This way, the model learns an inner representation of the English language that can then be used to extract features
36
+ useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard
37
+ classifier using the features produced by the MultiBERTs model as inputs.
38
 
39
  ## Intended uses & limitations
40
+
41
  You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
42
  be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=multiberts) to look for
43
  fine-tuned versions on a task that interests you.
 
46
  generation you should look at model like GPT2.
47
 
48
  ### How to use
49
+
50
  Here is how to use this model to get the features of a given text in PyTorch:
51
+
52
  ```python
53
  from transformers import BertTokenizer, BertModel
54
+ tokenizer = BertTokenizer.from_pretrained('multiberts-seed-11')
55
  model = BertModel.from_pretrained("multiberts-seed-11")
56
  text = "Replace me by any text you'd like."
57
  encoded_input = tokenizer(text, return_tensors='pt')
 
59
  ```
60
 
61
  ### Limitations and bias
62
+
63
  Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
64
  predictions. This bias will also affect all fine-tuned versions of this model. For an understanding of bias of this particular
65
  checkpoint, please try out this checkpoint with the snippet present in the [Limitation and bias section](https://huggingface.co/bert-base-uncased#limitations-and-bias) of the [bert-base-uncased](https://huggingface.co/bert-base-uncased) checkpoint.
66
 
67
  ## Training data
68
+
69
  The MultiBERTs models were pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038
70
  unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and
71
  headers).
72
+
73
  ## Training procedure
74
 
75
  ### Preprocessing
76
+
77
  The texts are lowercased and tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are
78
  then of the form:
79
+
80
  ```
81
  [CLS] Sentence A [SEP] Sentence B [SEP]
82
  ```
83
+
84
  With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in
85
  the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a
86
  consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two
87
  "sentences" has a combined length of less than 512 tokens.
88
  The details of the masking procedure for each sentence are the following:
89
+
90
  - 15% of the tokens are masked.
91
  - In 80% of the cases, the masked tokens are replaced by `[MASK]`.
92
  - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
93
  - In the 10% remaining cases, the masked tokens are left as is.
94
 
95
  ### Pretraining
96
+
97
  The model was trained on 16 Cloud TPU v2 chips for two million steps with a batch size
98
  of 256. The sequence length was set to 512 throughout. The optimizer
99
  used is Adam with a learning rate of 1e-4, \\(\beta_{1} = 0.9\\) and \\(\beta_{2} = 0.999\\), a weight decay of 0.01,
100
  learning rate warmup for 10,000 steps and linear decay of the learning rate after.
101
 
102
  ### BibTeX entry and citation info
103
+
104
  ```bibtex
105
  @article{DBLP:journals/corr/abs-2106-16163,
106
  author = {Thibault Sellam and
 
127
  bibsource = {dblp computer science bibliography, https://dblp.org}
128
  }
129
  ```
130
+
131
  <a href="https://huggingface.co/exbert/?model=multiberts">
132
  <img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
133
  </a>