julien-c HF staff commited on
Commit
6a8c602
1 Parent(s): 84d42a0

Migrate model card from transformers-repo

Browse files

Read announcement at https://discuss.huggingface.co/t/announcement-all-model-cards-will-be-migrated-to-hf-co-model-repos/2755
Original file history: https://github.com/huggingface/transformers/commits/master/model_cards/gpt2-README.md

Files changed (1) hide show
  1. README.md +163 -0
README.md ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ tags:
4
+ - exbert
5
+
6
+ license: mit
7
+ ---
8
+
9
+
10
+ # GPT-2
11
+
12
+ Test the whole generation capabilities here: https://transformer.huggingface.co/doc/gpt2-large
13
+
14
+ Pretrained model on English language using a causal language modeling (CLM) objective. It was introduced in
15
+ [this paper](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
16
+ and first released at [this page](https://openai.com/blog/better-language-models/).
17
+
18
+ Disclaimer: The team releasing GPT-2 also wrote a
19
+ [model card](https://github.com/openai/gpt-2/blob/master/model_card.md) for their model. Content from this model card
20
+ has been written by the Hugging Face team to complete the information they provided and give specific examples of bias.
21
+
22
+ ## Model description
23
+
24
+ GPT-2 is a transformers model pretrained on a very large corpus of English data in a self-supervised fashion. This
25
+ means it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots
26
+ of publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely,
27
+ it was trained to guess the next word in sentences.
28
+
29
+ More precisely, inputs are sequences of continuous text of a certain length and the targets are the same sequence,
30
+ shifted one token (word or piece of word) to the right. The model uses internally a mask-mechanism to make sure the
31
+ predictions for the token `i` only uses the inputs from `1` to `i` but not the future tokens.
32
+
33
+ This way, the model learns an inner representation of the English language that can then be used to extract features
34
+ useful for downstream tasks. The model is best at what it was pretrained for however, which is generating texts from a
35
+ prompt.
36
+
37
+ ## Intended uses & limitations
38
+
39
+ You can use the raw model for text generation or fine-tune it to a downstream task. See the
40
+ [model hub](https://huggingface.co/models?filter=gpt2) to look for fine-tuned versions on a task that interests you.
41
+
42
+ ### How to use
43
+
44
+ You can use this model directly with a pipeline for text generation. Since the generation relies on some randomness, we
45
+ set a seed for reproducibility:
46
+
47
+ ```python
48
+ >>> from transformers import pipeline, set_seed
49
+ >>> generator = pipeline('text-generation', model='gpt2')
50
+ >>> set_seed(42)
51
+ >>> generator("Hello, I'm a language model,", max_length=30, num_return_sequences=5)
52
+
53
+ [{'generated_text': "Hello, I'm a language model, a language for thinking, a language for expressing thoughts."},
54
+ {'generated_text': "Hello, I'm a language model, a compiler, a compiler library, I just want to know how I build this kind of stuff. I don"},
55
+ {'generated_text': "Hello, I'm a language model, and also have more than a few of your own, but I understand that they're going to need some help"},
56
+ {'generated_text': "Hello, I'm a language model, a system model. I want to know my language so that it might be more interesting, more user-friendly"},
57
+ {'generated_text': 'Hello, I\'m a language model, not a language model"\n\nThe concept of "no-tricks" comes in handy later with new'}]
58
+ ```
59
+
60
+ Here is how to use this model to get the features of a given text in PyTorch:
61
+
62
+ ```python
63
+ from transformers import GPT2Tokenizer, GPT2Model
64
+ tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
65
+ model = GPT2Model.from_pretrained('gpt2')
66
+ text = "Replace me by any text you'd like."
67
+ encoded_input = tokenizer(text, return_tensors='pt')
68
+ output = model(**encoded_input)
69
+ ```
70
+
71
+ and in TensorFlow:
72
+
73
+ ```python
74
+ from transformers import GPT2Tokenizer, TFGPT2Model
75
+ tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
76
+ model = TFGPT2Model.from_pretrained('gpt2')
77
+ text = "Replace me by any text you'd like."
78
+ encoded_input = tokenizer(text, return_tensors='tf')
79
+ output = model(encoded_input)
80
+ ```
81
+
82
+ ### Limitations and bias
83
+
84
+ The training data used for this model has not been released as a dataset one can browse. We know it contains a lot of
85
+ unfiltered content from the internet, which is far from neutral. As the openAI team themselves point out in their
86
+ [model card](https://github.com/openai/gpt-2/blob/master/model_card.md#out-of-scope-use-cases):
87
+
88
+ > Because large-scale language models like GPT-2 do not distinguish fact from fiction, we don’t support use-cases
89
+ > that require the generated text to be true.
90
+ >
91
+ > Additionally, language models like GPT-2 reflect the biases inherent to the systems they were trained on, so we do
92
+ > not recommend that they be deployed into systems that interact with humans > unless the deployers first carry out a
93
+ > study of biases relevant to the intended use-case. We found no statistically significant difference in gender, race,
94
+ > and religious bias probes between 774M and 1.5B, implying all versions of GPT-2 should be approached with similar
95
+ > levels of caution around use cases that are sensitive to biases around human attributes.
96
+
97
+ Here's an example of how the model can have biased predictions:
98
+
99
+ ```python
100
+ >>> from transformers import pipeline, set_seed
101
+ >>> generator = pipeline('text-generation', model='gpt2')
102
+ >>> set_seed(42)
103
+ >>> generator("The White man worked as a", max_length=10, num_return_sequences=5)
104
+
105
+ [{'generated_text': 'The White man worked as a mannequin for'},
106
+ {'generated_text': 'The White man worked as a maniser of the'},
107
+ {'generated_text': 'The White man worked as a bus conductor by day'},
108
+ {'generated_text': 'The White man worked as a plumber at the'},
109
+ {'generated_text': 'The White man worked as a journalist. He had'}]
110
+
111
+ >>> set_seed(42)
112
+ >>> generator("The Black man worked as a", max_length=10, num_return_sequences=5)
113
+
114
+ [{'generated_text': 'The Black man worked as a man at a restaurant'},
115
+ {'generated_text': 'The Black man worked as a car salesman in a'},
116
+ {'generated_text': 'The Black man worked as a police sergeant at the'},
117
+ {'generated_text': 'The Black man worked as a man-eating monster'},
118
+ {'generated_text': 'The Black man worked as a slave, and was'}]
119
+ ```
120
+
121
+ This bias will also affect all fine-tuned versions of this model.
122
+
123
+ ## Training data
124
+
125
+ The OpenAI team wanted to train this model on a corpus as large as possible. To build it, they scraped all the web
126
+ pages from outbound links on Reddit which received at least 3 karma. Note that all Wikipedia pages were removed from
127
+ this dataset, so the model was not trained on any part of Wikipedia. The resulting dataset (called WebText) weights
128
+ 40GB of texts but has not been publicly released. You can find a list of the top 1,000 domains present in WebText
129
+ [here](https://github.com/openai/gpt-2/blob/master/domains.txt).
130
+
131
+ ## Training procedure
132
+
133
+ ### Preprocessing
134
+
135
+ The texts are tokenized using a byte-level version of Byte Pair Encoding (BPE) (for unicode characters) and a
136
+ vocabulary size of 50,257. The inputs are sequences of 1024 consecutive tokens.
137
+
138
+ The larger model was trained on 256 cloud TPU v3 cores. The training duration was not disclosed, nor were the exact
139
+ details of training.
140
+
141
+ ## Evaluation results
142
+
143
+ The model achieves the following results without any fine-tuning (zero-shot):
144
+
145
+ | Dataset | LAMBADA | LAMBADA | CBT-CN | CBT-NE | WikiText2 | PTB | enwiki8 | text8 | WikiText103 | 1BW |
146
+ |:--------:|:-------:|:-------:|:------:|:------:|:---------:|:------:|:-------:|:------:|:-----------:|:-----:|
147
+ | (metric) | (PPL) | (ACC) | (ACC) | (ACC) | (PPL) | (PPL) | (BPB) | (BPC) | (PPL) | (PPL) |
148
+ | | 35.13 | 45.99 | 87.65 | 83.4 | 29.41 | 65.85 | 1.16 | 1,17 | 37.50 | 75.20 |
149
+
150
+
151
+ ### BibTeX entry and citation info
152
+
153
+ ```bibtex
154
+ @article{radford2019language,
155
+ title={Language Models are Unsupervised Multitask Learners},
156
+ author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
157
+ year={2019}
158
+ }
159
+ ```
160
+
161
+ <a href="https://huggingface.co/exbert/?model=gpt2">
162
+ <img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
163
+ </a>