FacebookAI/roberta-large-mnli

1---2language:3- en4license: mit5tags:6- autogenerated-modelcard7datasets:8- multi_nli9- wikipedia10- bookcorpus11---12 13# roberta-large-mnli14 15## Table of Contents16- [Model Details](#model-details)17- [How To Get Started With the Model](#how-to-get-started-with-the-model)18- [Uses](#uses)19- [Risks, Limitations and Biases](#risks-limitations-and-biases)20- [Training](#training)21- [Evaluation](#evaluation-results)22- [Environmental Impact](#environmental-impact)23- [Technical Specifications](#technical-specifications)24- [Citation Information](#citation-information)25- [Model Card Authors](#model-card-author)26 27## Model Details28 29**Model Description:** roberta-large-mnli is the [RoBERTa large model](https://huggingface.co/roberta-large) fine-tuned on the [Multi-Genre Natural Language Inference (MNLI)](https://huggingface.co/datasets/multi_nli) corpus. The model is a pretrained model on English language text using a masked language modeling (MLM) objective.30 31- **Developed by:** See [GitHub Repo](https://github.com/facebookresearch/fairseq/tree/main/examples/roberta) for model developers32- **Model Type:** Transformer-based language model33- **Language(s):** English34- **License:** MIT 35- **Parent Model:** This model is a fine-tuned version of the RoBERTa large model. Users should see the [RoBERTa large model card](https://huggingface.co/roberta-large) for relevant information.36- **Resources for more information:**37  - [Research Paper](https://arxiv.org/abs/1907.11692)38  - [GitHub Repo](https://github.com/facebookresearch/fairseq/tree/main/examples/roberta)39 40## How to Get Started with the Model 41 42Use the code below to get started with the model. The model can be loaded with the zero-shot-classification pipeline like so:43 44```python45from transformers import pipeline46classifier = pipeline('zero-shot-classification', model='roberta-large-mnli')47```48 49You can then use this pipeline to classify sequences into any of the class names you specify. For example:50 51```python52sequence_to_classify = "one day I will see the world"53candidate_labels = ['travel', 'cooking', 'dancing']54classifier(sequence_to_classify, candidate_labels)55```56 57## Uses58 59#### Direct Use60 61This fine-tuned model can be used for zero-shot classification tasks, including zero-shot sentence-pair classification (see the [GitHub repo](https://github.com/facebookresearch/fairseq/tree/main/examples/roberta) for examples) and zero-shot sequence classification.62 63#### Misuse and Out-of-scope Use64 65The model should not be used to intentionally create hostile or alienating environments for people. In addition, the model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.66 67## Risks, Limitations and Biases68 69**CONTENT WARNING: Readers should be aware this section contains content that is disturbing, offensive, and can propogate historical and current stereotypes.**70 71Significant research has explored bias and fairness issues with language models (see, e.g., [Sheng et al. (2021)](https://aclanthology.org/2021.acl-long.330.pdf) and [Bender et al. (2021)](https://dl.acm.org/doi/pdf/10.1145/3442188.3445922)). The [RoBERTa large model card](https://huggingface.co/roberta-large) notes that: "The training data used for this model contains a lot of unfiltered content from the internet, which is far from neutral." 72 73Predictions generated by the model can include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups. For example:74 75```python76sequence_to_classify = "The CEO had a strong handshake."77candidate_labels = ['male', 'female']78hypothesis_template = "This text speaks about a {} profession."79classifier(sequence_to_classify, candidate_labels, hypothesis_template=hypothesis_template)80```81 82Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.83 84## Training85 86#### Training Data87 88This model was fine-tuned on the [Multi-Genre Natural Language Inference (MNLI)](https://cims.nyu.edu/~sbowman/multinli/) corpus. Also see the [MNLI data card](https://huggingface.co/datasets/multi_nli) for more information. 89 90As described in the [RoBERTa large model card](https://huggingface.co/roberta-large): 91 92> The RoBERTa model was pretrained on the reunion of five datasets:93> 94> - [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038 unpublished books;95> - [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and headers) ;96> - [CC-News](https://commoncrawl.org/2016/10/news-dataset-available/), a dataset containing 63 millions English news articles crawled between September 2016 and February 2019.97> - [OpenWebText](https://github.com/jcpeterson/openwebtext), an opensource recreation of the WebText dataset used to train GPT-2,98> - [Stories](https://arxiv.org/abs/1806.02847), a dataset containing a subset of CommonCrawl data filtered to match the story-like style of Winograd schemas.99>100> Together theses datasets weight 160GB of text.101 102Also see the [bookcorpus data card](https://huggingface.co/datasets/bookcorpus) and the [wikipedia data card](https://huggingface.co/datasets/wikipedia) for additional information.103 104#### Training Procedure105 106##### Preprocessing107 108As described in the [RoBERTa large model card](https://huggingface.co/roberta-large): 109 110> The texts are tokenized using a byte version of Byte-Pair Encoding (BPE) and a vocabulary size of 50,000. The inputs of111> the model take pieces of 512 contiguous token that may span over documents. The beginning of a new document is marked112> with `<s>` and the end of one by `</s>`113> 114> The details of the masking procedure for each sentence are the following:115> - 15% of the tokens are masked.116> - In 80% of the cases, the masked tokens are replaced by `<mask>`.117> - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.118> - In the 10% remaining cases, the masked tokens are left as is.119> 120> Contrary to BERT, the masking is done dynamically during pretraining (e.g., it changes at each epoch and is not fixed).121 122##### Pretraining 123 124Also as described in the [RoBERTa large model card](https://huggingface.co/roberta-large): 125 126> The model was trained on 1024 V100 GPUs for 500K steps with a batch size of 8K and a sequence length of 512. The127> optimizer used is Adam with a learning rate of 4e-4, \\(\beta_{1} = 0.9\\), \\(\beta_{2} = 0.98\\) and128> \\(\epsilon = 1e-6\\), a weight decay of 0.01, learning rate warmup for 30,000 steps and linear decay of the learning129> rate after.130 131## Evaluation132 133The following evaluation information is extracted from the associated [GitHub repo for RoBERTa](https://github.com/facebookresearch/fairseq/tree/main/examples/roberta). 134 135#### Testing Data, Factors and Metrics136 137The model developers report that the model was evaluated on the following tasks and datasets using the listed metrics: 138 139- **Dataset:** Part of [GLUE (Wang et al., 2019)](https://arxiv.org/pdf/1804.07461.pdf), the General Language Understanding Evaluation benchmark, a collection of 9 datasets for evaluating natural language understanding systems. Specifically, the model was evaluated on the [Multi-Genre Natural Language Inference (MNLI)](https://cims.nyu.edu/~sbowman/multinli/) corpus. See the [GLUE data card](https://huggingface.co/datasets/glue) or [Wang et al. (2019)](https://arxiv.org/pdf/1804.07461.pdf) for further information.140  - **Tasks:** NLI. [Wang et al. (2019)](https://arxiv.org/pdf/1804.07461.pdf) describe the inference task for MNLI as: 141  > The Multi-Genre Natural Language Inference Corpus [(Williams et al., 2018)](https://arxiv.org/abs/1704.05426) is a crowd-sourced collection of sentence pairs with textual entailment annotations. Given a premise sentence and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis (entailment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are gathered from ten different sources, including transcribed speech, fiction, and government reports. We use the standard test set, for which we obtained private labels from the authors, and evaluate on both the matched (in-domain) and mismatched (cross-domain) sections. We also use and recommend the SNLI corpus [(Bowman et al., 2015)](https://arxiv.org/abs/1508.05326) as 550k examples of auxiliary training data.142  - **Metrics:** Accuracy  143  144- **Dataset:** [XNLI (Conneau et al., 2018)](https://arxiv.org/pdf/1809.05053.pdf), the extension of the [Multi-Genre Natural Language Inference (MNLI)](https://cims.nyu.edu/~sbowman/multinli/) corpus to 15 languages: English, French, Spanish, German, Greek, Bulgarian, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, Hindi, Swahili and Urdu. See the [XNLI data card](https://huggingface.co/datasets/xnli) or [Conneau et al. (2018)](https://arxiv.org/pdf/1809.05053.pdf) for further information.145  - **Tasks:** Translate-test (e.g., the model is used to translate input sentences in other languages to the training language)146  - **Metrics:** Accuracy147 148#### Results149 150GLUE test results (dev set, single model, single-task fine-tuning): 90.2 on MNLI151 152XNLI test results:153 154| Task | en |  fr | es  | de  | el  | bg  | ru  | tr  | ar  | vi  | th  | zh  | hi  | sw  | ur  |155|:----:|:--:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|156|      |91.3|82.91|84.27|81.24|81.74|83.13|78.28|76.79|76.64|74.17|74.05| 77.5| 70.9|66.65|66.81|	157 158## Environmental Impact159 160Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). We present the hardware type and hours used based on the [associated paper](https://arxiv.org/pdf/1907.11692.pdf).161 162- **Hardware Type:** 1024 V100 GPUs163- **Hours used:** 24 hours (one day)164- **Cloud Provider:** Unknown165- **Compute Region:** Unknown166- **Carbon Emitted:** Unknown167 168## Technical Specifications169 170See the [associated paper](https://arxiv.org/pdf/1907.11692.pdf) for details on the modeling architecture, objective, compute infrastructure, and training details.171 172## Citation Information173 174```bibtex175@article{liu2019roberta,176    title = {RoBERTa: A Robustly Optimized BERT Pretraining Approach},177    author = {Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and178              Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and179              Luke Zettlemoyer and Veselin Stoyanov},180    journal={arXiv preprint arXiv:1907.11692},181    year = {2019},182}183```