README.md

---
language:
- en
metrics:
- accuracy
pipeline_tag: fill-mask
tags:
- not-for-all-audiences
- abusive language
- hate speech
- offensive language
widget:
- text: They is a [MASK].
  example_title: Neutral
- text: She is a [MASK].
  example_title: Misogyny
- text: He is a [MASK].
  example_title: Misandry
---

**WARNING: Some language produced by this model and README may offend. The model intent is to facilitate bias in AI research**

# notSexistBERT base model (uncased)

Re-pretrained model on English language using a Masked Language Modeling (MLM)
and Next Sentence Prediction (NSP) objective. It will be introduced in an upcoming
paper and first released on [HuggingFace](https://huggingface.co/clincolnoz/notSexistBERT_temp). This model is uncased: it does not make a difference between english and English.

## Model description

sexistBERT is a transformers model pretrained on a **less sexist** corpus of English data in a
self-supervised fashion. This means it was pretrained on the raw texts only,
with no humans labeling them in any way (which is why it can use lots of
publicly available data) with an automatic process to generate inputs and labels
from those texts. More precisely, it was pretrained with two objectives:

- Masked language modeling (MLM): taking a sentence, the model randomly masks
  15% of the words in the input then run the entire masked sentence through the
  model and has to predict the masked words. This is different from traditional
  recurrent neural networks (RNNs) that usually see the words one after the
  other, or from autoregressive models like GPT which internally masks the
  future tokens. It allows the model to learn a bidirectional representation of
  the sentence.
- Next sentence prediction (NSP): the models concatenates two masked sentences
  as inputs during pretraining. Sometimes they correspond to sentences that were
  next to each other in the original text, sometimes not. The model then has to
  predict if the two sentences were following each other or not.

This way, the model learns an inner representation of the English language that
can then be used to extract features useful for downstream tasks: if you have a
dataset of labeled sentences, for instance, you can train a standard classifier
using the features produced by the BERT model as inputs.

## Model variations

notSexistBERT has originally been released as sexist and notSexist variations. The uncased models strip out any accent markers.

| Model                                                                   | #params   | Language |
| ----------------------------------------------------------------------- | --------- | -------- |
| [`sexistBERT`](https://huggingface.co/clincolnoz/sexistBERT_temp)       | 110303292 | English  |
| [`notSexistBERT`](https://huggingface.co/clincolnoz/notSexistBERT_temp) | 110201784 | English  |

## Intended uses & limitations

Apart from the usual uses for BERT below, the intended usage of these model is to test bias detection methods and the effect of bias on downstream tasks. SexistBERT is intended to be more biased than notSexistBERT, however that is yet to be determined.

You can use the raw model for either masked language modeling or next sentence
prediction, but it's mostly intended to be fine-tuned on a downstream task. See
the [model hub](https://huggingface.co/models?filter=bert) to look for
fine-tuned versions of a task that interests you.

Note that this model is primarily aimed at being fine-tuned on tasks that use
the whole sentence (potentially masked) to make decisions, such as sequence
classification, token classification or question answering.

For tasks such as text generation you should look at model like GPT2.

### How to use

You can use this model directly with a pipeline for masked language modeling:

```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='clincolnoz/notSexistBERT_temp')
>>> unmasker("Hello I'm a [MASK] model.")

[{'score': 0.5223352313041687,
  'token': 2535,
  'token_str': 'role',
  'sequence': "hello i'm a role model."},
 {'score': 0.12853220105171204,
  'token': 2449,
  'token_str': 'business',
  'sequence': "hello i'm a business model."},
 {'score': 0.0621086061000824,
  'token': 3287,
  'token_str': 'male',
  'sequence': "hello i'm a male model."},
 {'score': 0.03042026236653328,
  'token': 3565,
  'token_str': 'super',
  'sequence': "hello i'm a super model."},
 {'score': 0.01949389837682247,
  'token': 7605,
  'token_str': '3d',
  'sequence': "hello i'm a 3d model."}]
```

Here is how to use this model to get the features of a given text in PyTorch:

```python
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained(
  'clincolnoz/notSexistBERT_temp',
  revision='v0.34' # tag name, or branch name, or commit hash
)
model = BertModel.from_pretrained(
  'clincolnoz/notSexistBERT_temp',
  revision='v0.34' # tag name, or branch name, or commit hash
)
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```

and in TensorFlow:

```python
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained(
  'clincolnoz/notSexistBERT_temp',
  revision='v0.34' # tag name, or branch name, or commit hash
)
model = TFBertModel.from_pretrained(
  'clincolnoz/notSexistBERT_temp',
  from_pt=True,
  revision='v0.34' # tag name, or branch name, or commit hash
)
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```

### Limitations and bias

Even if the training data used for this model could be characterized as fairly
neutral, this model can have biased predictions:

```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='clincolnoz/notSexistBERT_temp')
>>> unmasker("The man worked as a [MASK].")

[{'score': 0.1064024269580841,
  'token': 5160,
  'token_str': 'lawyer',
  'sequence': 'the man worked as a lawyer.'},
 {'score': 0.06261951476335526,
  'token': 7155,
  'token_str': 'scientist',
  'sequence': 'the man worked as a scientist.'},
 {'score': 0.046040475368499756,
  'token': 10563,
  'token_str': 'teenager',
  'sequence': 'the man worked as a teenager.'},
 {'score': 0.04330913722515106,
  'token': 20273,
  'token_str': 'programmer',
  'sequence': 'the man worked as a programmer.'},
 {'score': 0.04167287424206734,
  'token': 5766,
  'token_str': 'ceo',
  'sequence': 'the man worked as a ceo.'}]

>>> unmasker("The woman worked as a [MASK].")

[{'score': 0.0949002057313919,
  'token': 6821,
  'token_str': 'nurse',
  'sequence': 'the woman worked as a nurse.'},
 {'score': 0.08425672352313995,
  'token': 3208,
  'token_str': 'manager',
  'sequence': 'the woman worked as a manager.'},
 {'score': 0.07672832906246185,
  'token': 5160,
  'token_str': 'lawyer',
  'sequence': 'the woman worked as a lawyer.'},
 {'score': 0.042527567595243454,
  'token': 7522,
  'token_str': 'physician',
  'sequence': 'the woman worked as a physician.'},
 {'score': 0.034959811717271805,
  'token': 5766,
  'token_str': 'ceo',
  'sequence': 'the woman worked as a ceo.'}]
```

This bias may also affect all fine-tuned versions of this model.

## Training data

TBD

<!-- The BERT model was pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038
unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and
headers). -->

## Training procedure

### Preprocessing

For the NSP task the data were preprocessed by splitting documents into sentences to create first a bag of sentences and then to create pairs of sentences, where Sentence B either corresponded to a consecutive sentence in the text or randomly select from the bag. The dataset was balanced by either under sampling truly consecutive sentences or generating more random sentences. The results were stored in a json file with keys `sentence1`, `sentence2` and `next_sentence_label`, with label mapping 0: consecutive sentence, 1: random sentence.

The texts are lowercased and tokenized using WordPiece and a vocabulary size of
30,646. The inputs of the model are then of the form:

```
[CLS] Sentence A [SEP] Sentence B [SEP]
```

With probability 0.5, sentence A and sentence B correspond to two consecutive
sentences in the original corpus, and in the other cases, it's another random
sentence in the corpus. Note that what is considered a sentence here is a
consecutive span of text usually longer than a single sentence. The only
constrain is that the result with the two "sentences" has a combined length of
less than 512 tokens.

The details of the masking procedure for each sentence are the following:

- 15% of the tokens are masked.
- In 80% of the cases, the masked tokens are replaced by `[MASK]`.
- In 10% of the cases, the masked tokens are replaced by a random token
  (different) from the one they replace.
- In the 10% remaining cases, the masked tokens are left as is.

### Pretraining

The model was trained on a NVIDIA GeForce RTX 4090 using 16-bit precision for 20
million steps with a batch size of 24. The sequence length was limited 512. The
optimizer used is Adam with a learning rate of 5e-5, \\(\beta*{1} = 0.9\\) and
\\(\beta*{2} = 0.999\\), a weight decay of 0.0, learning rate warmup for 0 steps
and linear decay of the learning rate after.

<!-- ## Evaluation results

When fine-tuned on downstream tasks, this model achieves the following results:

Glue test results:

| Task  | MNLI-(m/mm) |  QQP  | QNLI  | SST-2 | CoLA  | STS-B | MRPC  |  RTE  | Average |
| :---: | :---------: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :-----: |
|       |  84.6/83.4  | 71.2  | 90.5  | 93.5  | 52.1  | 85.8  | 88.9  | 66.4  |  79.6   | --> |

### Framework versions

- Transformers 4.27.0.dev0
- Pytorch 1.13.1+cu117
- Datasets 2.9.0
- Tokenizers 0.13.2

<!-- ### BibTeX entry and citation info -->

<!-- ```bibtex
@article{
  
}
``` -->