Commit
•
f4df4fa
1
Parent(s):
243d3e1
Add training language statistics
Browse files
README.md
CHANGED
|
@@ -8,6 +8,43 @@ inference: false
|
|
| 8 |
|
| 9 |
XGLM-2.9B is a multilingual autoregressive language model (with 2.9 billion parameters) trained on a balanced corpus of a diverse set of languages totaling 500 billion sub-tokens. It was introduced in the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin\*, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li\* (\*Equal Contribution). The original implementation was released in [this repository](https://github.com/pytorch/fairseq/tree/main/examples/xglm).
|
| 10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
## Model card
|
| 12 |
|
| 13 |
For intended usage of the model, please refer to the [model card](https://github.com/pytorch/fairseq/blob/main/examples/xglm/model_card.md) released by the team releasing XGLM-2.9B.
|
|
|
|
| 8 |
|
| 9 |
XGLM-2.9B is a multilingual autoregressive language model (with 2.9 billion parameters) trained on a balanced corpus of a diverse set of languages totaling 500 billion sub-tokens. It was introduced in the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin\*, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li\* (\*Equal Contribution). The original implementation was released in [this repository](https://github.com/pytorch/fairseq/tree/main/examples/xglm).
|
| 10 |
|
| 11 |
+
## Training Data Statistics
|
| 12 |
+
|
| 13 |
+
The training data statistics of XGLM-2.9B is shown in the table below.
|
| 14 |
+
|
| 15 |
+
| ISO-639-1| family | name | # tokens | ratio | ratio w/ lowRes upsampling |
|
| 16 |
+
|:--------|:-----------------|:------------------------|-------------:|------------:|-------------:|
|
| 17 |
+
| en | Indo-European | English | 803526736124 | 0.489906 | 0.3259 |
|
| 18 |
+
| ru | Indo-European | Russian | 147791898098 | 0.0901079 | 0.0602 |
|
| 19 |
+
| zh | Sino-Tibetan | Chinese | 132770494630 | 0.0809494 | 0.0483 |
|
| 20 |
+
| de | Indo-European | German | 89223707856 | 0.0543992 | 0.0363 |
|
| 21 |
+
| es | Indo-European | Spanish | 87303083105 | 0.0532282 | 0.0353 |
|
| 22 |
+
| fr | Indo-European | French | 77419639775 | 0.0472023 | 0.0313 |
|
| 23 |
+
| ja | Japonic | Japanese | 66054364513 | 0.040273 | 0.0269 |
|
| 24 |
+
| it | Indo-European | Italian | 41930465338 | 0.0255648 | 0.0171 |
|
| 25 |
+
| pt | Indo-European | Portuguese | 36586032444 | 0.0223063 | 0.0297 |
|
| 26 |
+
| el | Indo-European | Greek (modern) | 28762166159 | 0.0175361 | 0.0233 |
|
| 27 |
+
| ko | Koreanic | Korean | 20002244535 | 0.0121953 | 0.0811 |
|
| 28 |
+
| fi | Uralic | Finnish | 16804309722 | 0.0102455 | 0.0681 |
|
| 29 |
+
| id | Austronesian | Indonesian | 15423541953 | 0.00940365 | 0.0125 |
|
| 30 |
+
| tr | Turkic | Turkish | 12413166065 | 0.00756824 | 0.0101 |
|
| 31 |
+
| ar | Afro-Asiatic | Arabic | 12248607345 | 0.00746791 | 0.0099 |
|
| 32 |
+
| vi | Austroasiatic | Vietnamese | 11199121869 | 0.00682804 | 0.0091 |
|
| 33 |
+
| th | Tai–Kadai | Thai | 10842172807 | 0.00661041 | 0.044 |
|
| 34 |
+
| bg | Indo-European | Bulgarian | 9703797869 | 0.00591635 | 0.0393 |
|
| 35 |
+
| ca | Indo-European | Catalan | 7075834775 | 0.0043141 | 0.0287 |
|
| 36 |
+
| hi | Indo-European | Hindi | 3448390110 | 0.00210246 | 0.014 |
|
| 37 |
+
| et | Uralic | Estonian | 3286873851 | 0.00200399 | 0.0133 |
|
| 38 |
+
| bn | Indo-European | Bengali, Bangla | 1627447450 | 0.000992245 | 0.0066 |
|
| 39 |
+
| ta | Dravidian | Tamil | 1476973397 | 0.000900502 | 0.006 |
|
| 40 |
+
| ur | Indo-European | Urdu | 1351891969 | 0.000824241 | 0.0055 |
|
| 41 |
+
| sw | Niger–Congo | Swahili | 907516139 | 0.000553307 | 0.0037 |
|
| 42 |
+
| te | Dravidian | Telugu | 689316485 | 0.000420272 | 0.0028 |
|
| 43 |
+
| eu | Language isolate | Basque | 105304423 | 6.42035e-05 | 0.0043 |
|
| 44 |
+
| my | Sino-Tibetan | Burmese | 101358331 | 6.17976e-05 | 0.003 |
|
| 45 |
+
| ht | Creole | Haitian, Haitian Creole | 86584697 | 5.27902e-05 | 0.0035 |
|
| 46 |
+
| qu | Quechuan | Quechua | 3236108 | 1.97304e-06 | 0.0001 |
|
| 47 |
+
|
| 48 |
## Model card
|
| 49 |
|
| 50 |
For intended usage of the model, please refer to the [model card](https://github.com/pytorch/fairseq/blob/main/examples/xglm/model_card.md) released by the team releasing XGLM-2.9B.
|