Datasets:

ArXiv:
akshitab commited on
Commit
fa00680
1 Parent(s): 540b94c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +11 -8
README.md CHANGED
@@ -52,7 +52,7 @@ git clone https://huggingface.co/datasets/allenai/nllb
52
 
53
  ### Supported Tasks and Leaderboards
54
 
55
- NA
56
 
57
  ### Languages
58
 
@@ -81,6 +81,8 @@ Every instance for a language pair contains the following fields: 'translation'
81
  * Second sentence source
82
  * Second sentence URL if the source is crawl-data/\*; _ otherwise
83
 
 
 
84
  Example:
85
  ```
86
  {'translation': {'ace_Latn': 'Gobnyan hana geupeukeucewa gata atawa geutinggai meunan mantong gata."',
@@ -96,14 +98,14 @@ Example:
96
 
97
  ### Data Splits
98
 
99
- The data is not split. Given the noisy nature of the overall process, we recommend using the data only for training and use other datasets like [Flores-200](https://github.com/facebookresearch/flores) for the evaluation.
100
 
101
 
102
  ## Dataset Creation
103
 
104
  ### Curation Rationale
105
 
106
- Data was filtered based on language identification, emoji based filtering, and for some high-resource languages language model-based filtering. For more details on data filtering please refer to Section 5.2 (NLLB Team et al., 2022).
107
 
108
 
109
  ### Source Data
@@ -152,7 +154,7 @@ The data was not human annotated.
152
 
153
  ### Personal and Sensitive Information
154
 
155
- Data may contain personally identifiable information, sensitive or toxic content that was publicly shared on the Internet.
156
 
157
  ## Considerations for Using the Data
158
 
@@ -162,11 +164,11 @@ This dataset provides data for training machine learning systems for many langua
162
 
163
  ### Discussion of Biases
164
 
165
- Biases in the data have not been specifically studied, however as the original source of data is World Wide Web it is likely that the data has biases similar to those prevalent in the Internet. The data may also exhibit biases introduced by language identification and data filtering techniques: lower resource languages may have lower accuracy while data filtering techniques may remove certain less natural utterances.
166
 
167
  ### Other Known Limitations
168
 
169
- Some of the translations are in fact machine translation. Indeed, some sites have evidence of WordPress translation plugins in their HTML source. These sites were not filtered out en mass because HTML source was not available in many cases.
170
 
171
  ## Additional Information
172
 
@@ -181,8 +183,9 @@ The dataset is released under the terms of [ODC-BY](https://opendatacommons.org/
181
 
182
  ### Citation Information
183
 
184
- NLLB Team et al, No Language Left Behind: Scaling Human-Centered Machine Translation, Arxiv, 2022.
 
185
 
186
  ### Contributions
187
 
188
- We thank the AllenNLP team at AI2 for hosting and releasing this data, including [Akshita Bhagia](https://akshitab.github.io/) (for engineering efforts to create the huggingface dataset), and [Jesse Dodge](https://jessedodge.github.io/) (for organizing the connection).
 
52
 
53
  ### Supported Tasks and Leaderboards
54
 
55
+ N/A
56
 
57
  ### Languages
58
 
 
81
  * Second sentence source
82
  * Second sentence URL if the source is crawl-data/\*; _ otherwise
83
 
84
+ The lines are sorted by LASER3 score in decreasing order.
85
+
86
  Example:
87
  ```
88
  {'translation': {'ace_Latn': 'Gobnyan hana geupeukeucewa gata atawa geutinggai meunan mantong gata."',
 
98
 
99
  ### Data Splits
100
 
101
+ The data is not split. Given the noisy nature of the overall process, we recommend using the data only for training and use other datasets like [Flores-200](https://github.com/facebookresearch/flores) for the evaluation. The data includes some development and test sets from other datasets, such as xlsum. In addition, sourcing data from multiple web crawls is likely to produce incidental overlap with other test sets.
102
 
103
 
104
  ## Dataset Creation
105
 
106
  ### Curation Rationale
107
 
108
+ Data was filtered based on language identification, emoji based filtering, and for some high-resource languages using a language model. For more details on data filtering please refer to Section 5.2 (NLLB Team et al., 2022).
109
 
110
 
111
  ### Source Data
 
154
 
155
  ### Personal and Sensitive Information
156
 
157
+ Data may contain personally identifiable information, sensitive content, or toxic content that was publicly shared on the Internet.
158
 
159
  ## Considerations for Using the Data
160
 
 
164
 
165
  ### Discussion of Biases
166
 
167
+ Biases in the data have not been specifically studied, however as the original source of data is World Wide Web it is likely that the data has biases similar to those prevalent in the Internet. The data may also exhibit biases introduced by language identification and data filtering techniques; lower resource languages generally have lower accuracy.
168
 
169
  ### Other Known Limitations
170
 
171
+ Some of the translations are in fact machine translations. While some website machine translation tools are identifiable from HTML source, these tools were not filtered out en mass because raw HTML was not available from some sources and CommonCrawl processing started from WET files.
172
 
173
  ## Additional Information
174
 
 
183
 
184
  ### Citation Information
185
 
186
+ Hefferman et al, Bitext Mining Using Distilled Sentence Representations for Low-Resource Languages. Arxiv https://arxiv.org/abs/2205.12654, 2022.
187
+ NLLB Team et al, No Language Left Behind: Scaling Human-Centered Machine Translation, Arxiv https://arxiv.org/abs/2207.04672, 2022
188
 
189
  ### Contributions
190
 
191
+ We thank the NLLB Meta AI team for open sourcing the meta data and instructions on how to use it with special thanks to Bapi Akula, Pierre Andrews, Onur Çelebi, Sergey Edunov, Kenneth Heafield, Philipp Koehn, Alex Mourachko, Safiyyah Saleem, Holger Schwenk, and Guillaume Wenzek. We also thank the AllenNLP team at AI2 for hosting and releasing this data, including Akshita Bhagia (for engineering efforts to host the data, and create the huggingface dataset), and Jesse Dodge (for organizing the connection).