|
--- |
|
language: |
|
- multilingual |
|
- en |
|
- fr |
|
- es |
|
- de |
|
- it |
|
- pt |
|
- nl |
|
- sv |
|
- pl |
|
- ru |
|
- ar |
|
- tr |
|
- zh |
|
- ja |
|
- ko |
|
- hi |
|
- vi |
|
license: cc-by-nc-4.0 |
|
--- |
|
|
|
# xlm-mlm-17-1280 |
|
|
|
# Table of Contents |
|
|
|
1. [Model Details](#model-details) |
|
2. [Uses](#uses) |
|
3. [Bias, Risks, and Limitations](#bias-risks-and-limitations) |
|
4. [Training](#training) |
|
5. [Evaluation](#evaluation) |
|
6. [Environmental Impact](#environmental-impact) |
|
7. [Technical Specifications](#technical-specifications) |
|
8. [Citation](#citation) |
|
9. [Model Card Authors](#model-card-authors) |
|
10. [How To Get Started With the Model](#how-to-get-started-with-the-model) |
|
|
|
|
|
# Model Details |
|
|
|
xlm-mlm-17-1280 is the XLM model, which was proposed in [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau, trained on text in 17 languages. The model is a transformer pretrained using a masked language modeling (MLM) objective. |
|
|
|
## Model Description |
|
|
|
- **Developed by:** See [associated paper](https://arxiv.org/abs/1901.07291) and [GitHub Repo](https://github.com/facebookresearch/XLM) |
|
- **Model type:** Language model |
|
- **Language(s) (NLP):** 17 languages, see [GitHub Repo](https://github.com/facebookresearch/XLM#the-17-and-100-languages) for full list. |
|
- **License:** CC-BY-NC-4.0 |
|
- **Related Models:** [xlm-mlm-17-1280](https://huggingface.co/xlm-mlm-17-1280) |
|
- **Resources for more information:** |
|
- [Associated paper](https://arxiv.org/abs/1901.07291) |
|
- [GitHub Repo](https://github.com/facebookresearch/XLM#the-17-and-100-languages) |
|
- [Hugging Face Multilingual Models for Inference docs](https://huggingface.co/docs/transformers/v4.20.1/en/multilingual#xlm-with-language-embeddings) |
|
|
|
# Uses |
|
|
|
## Direct Use |
|
|
|
The model is a language model. The model can be used for masked language modeling. |
|
|
|
## Downstream Use |
|
|
|
To learn more about this task and potential downstream uses, see the Hugging Face [fill mask docs](https://huggingface.co/tasks/fill-mask) and the [Hugging Face Multilingual Models for Inference](https://huggingface.co/docs/transformers/v4.20.1/en/multilingual#xlm-with-language-embeddings) docs. Also see the [associated paper](https://arxiv.org/abs/1901.07291). |
|
|
|
## Out-of-Scope Use |
|
|
|
The model should not be used to intentionally create hostile or alienating environments for people. |
|
|
|
# Bias, Risks, and Limitations |
|
|
|
Significant 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)). |
|
|
|
## Recommendations |
|
|
|
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. |
|
|
|
# Training |
|
|
|
This model is the XLM model trained on text in 17 languages. The preprocessing included tokenization and byte-pair-encoding. See the [GitHub repo](https://github.com/facebookresearch/XLM#the-17-and-100-languages) and the [associated paper](https://arxiv.org/pdf/1911.02116.pdf) for further details on the training data and training procedure. |
|
|
|
[Conneau et al. (2020)](https://arxiv.org/pdf/1911.02116.pdf) report that this model has 16 layers, 1280 hidden states, 16 attention heads, and the dimension of the feed-forward layer is 1520. The vocabulary size is 200k and the total number of parameters is 570M (see Table 7). |
|
|
|
# Evaluation |
|
|
|
## Testing Data, Factors & Metrics |
|
|
|
The model developers evaluated the model on the XNLI cross-lingual classification task (see the [XNLI data card](https://huggingface.co/datasets/xnli) for more details on XNLI) using the metric of test accuracy. See the [GitHub Repo](https://arxiv.org/pdf/1911.02116.pdf) for further details on the testing data, factors and metrics. |
|
|
|
## Results |
|
|
|
For xlm-mlm-17-1280, the test accuracy on the XNLI cross-lingual classification task in English (en), Spanish (es), German (de), Arabic (ar), and Chinese (zh): |
|
|
|
|Language| en | es | de | ar | zh | |
|
|:------:|:--:|:---:|:--:|:--:|:--:| |
|
| |84.8|79.4 |76.2|71.5|75 | |
|
|
|
See the [GitHub repo](https://github.com/facebookresearch/XLM#ii-cross-lingual-language-model-pretraining-xlm) for further details. |
|
|
|
# Environmental Impact |
|
|
|
Carbon 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). |
|
|
|
- **Hardware Type:** More information needed |
|
- **Hours used:** More information needed |
|
- **Cloud Provider:** More information needed |
|
- **Compute Region:** More information needed |
|
- **Carbon Emitted:** More information needed |
|
|
|
# Technical Specifications |
|
|
|
[Conneau et al. (2020)](https://arxiv.org/pdf/1911.02116.pdf) report that this model has 16 layers, 1280 hidden states, 16 attention heads, and the dimension of the feed-forward layer is 1520. The vocabulary size is 200k and the total number of parameters is 570M (see Table 7). |
|
|
|
# Citation |
|
|
|
**BibTeX:** |
|
|
|
```bibtex |
|
@article{lample2019cross, |
|
title={Cross-lingual language model pretraining}, |
|
author={Lample, Guillaume and Conneau, Alexis}, |
|
journal={arXiv preprint arXiv:1901.07291}, |
|
year={2019} |
|
} |
|
``` |
|
|
|
**APA:** |
|
- Lample, G., & Conneau, A. (2019). Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291. |
|
|
|
# Model Card Authors |
|
|
|
This model card was written by the team at Hugging Face. |
|
|
|
# How to Get Started with the Model |
|
|
|
More information needed. See the [ipython notebook](https://github.com/facebookresearch/XLM/blob/main/generate-embeddings.ipynb) in the associated [GitHub repo](https://github.com/facebookresearch/XLM#the-17-and-100-languages) for examples. |
