Update README.md

README.md

@@ -43,10 +43,8 @@ by Meta AI and pretrained on the [CC100 multilingual dataset](https://huggingfac
 It was then fine-tuned on the [XNLI dataset](https://huggingface.co/datasets/xnli), which contains hypothesis-premise pairs from 15 languages, 
 as well as the English [MNLI dataset](https://huggingface.co/datasets/multi_nli).
 XLM-V-base was publish on 23.01.2023 in [this paper](https://arxiv.org/pdf/2301.10472.pdf). 
-Its main innovation is a larger vocabulary: previous multilingual models had a vocabulary of 250 000 tokens, 
-while XLM-V has 1 million tokens. The improved vocabulary allows for better representations of more languages. 
-
-[mDeBERTa-v3](https://arxiv.org/pdf/2111.09543.pdf). 
+Its main innovation is a larger and better vocabulary: previous multilingual models had a vocabulary of 250 000 tokens, 
+while XLM-V 'knows' 1 million tokens. The improved vocabulary allows for better representations of more languages. 
 
 
 ### How to use the model
@@ -89,7 +87,7 @@ Note that the XNLI contains a training set of 15 machine translated versions of
 but due to quality issues with these machine translations, this model was only trained on the professional translations 
 from the XNLI development set and the original English MNLI training set (392 702 texts). 
 Not using machine translated texts can avoid overfitting the model to the 15 languages; 
-avoids catastrophic forgetting of the other 85 languages mDeBERTa was pre-trained on; 
+avoids catastrophic forgetting of the other 101~ languages XLM-V was pre-trained on; 
 and significantly reduces training costs. 
 
 ### Training procedure
@@ -104,6 +102,7 @@ training_args = TrainingArguments(
     weight_decay=0.01,               # strength of weight decay
 )
 ```
+
 ### Eval results
 The model was evaluated on the XNLI test set on 15 languages (5010 texts per language, 75150 in total). 
 Note that multilingual NLI models are capable of classifying NLI texts without receiving NLI training data 
@@ -111,25 +110,36 @@ in the specific language (cross-lingual transfer). This means that the model is
 the other 101~ languages XLM-V was training on, but performance is most likely lower than for those languages available in XNLI.
 
 Also note that if other multilingual models on the model hub claim performance of around 90% on languages other than English, 
-the authors have most likely made a mistake during testing since non of the latest papers shows a multilingual average performance 
+the authors have most likely made a mistake during testing since non of the latest papers (of mostly larger models) shows a multilingual average performance 
 of more than a few points above 80% on XNLI (see [here](https://arxiv.org/pdf/2111.09543.pdf) or [here](https://arxiv.org/pdf/1911.02116.pdf)). 
 
-average | ar | bg | de | el | en | es | fr | hi | ru | sw | th | tr | ur | vi | zh 
----------|----------|---------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------
-0.808 | 0.802 | 0.829 | 0.825 | 0.826 | 0.883 | 0.845 | 0.834 | 0.771 | 0.813 | 0.748 | 0.793 | 0.807 | 0.740 | 0.795 | 0.8116
+The average XNLI performance of XLM-V reported in the paper is 0.76 ([see table 2](https://arxiv.org/pdf/2301.10472.pdf)). 
+This reimplementation has an average performance of 0.78. 
+This increase in performance is probably thanks to the addition of MNLI in the training data. 
+Note that [mDeBERTa-v3-base-mnli-xnli](https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-mnli-xnli) has an average 
+performance of 0.808 and is smaller (3GB for XLM-V vs. 560MB for mDeBERTa) and is faster (thanks to mDeBERTa's smaller vocabulary). 
+This difference comes probably from mDeBERTa-v3's improved pre-training objective. 
+Depending on the task, it is probably better to use [mDeBERTa-v3-base-mnli-xnli](https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-mnli-xnli),
+but XLM-V could be better on some languages based on its improved vocabulary.
 
+|Datasets|ar|bg|de|el|en|es|fr|hi|ru|sw|th|tr|ur|vi|zh|average
+| :---:  | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |:---: |
+|Accuracy|0.757|0.808|0.796|0.79|0.856|0.814|0.806|0.751|0.782|0.725|0.757|0.766|0.729|0.784|0.782|0.780|
+|Speed (text/sec)|3501.0|3324.0|3438.0|3174.0|3713.0|3500.0|3129.0|3042.0|3419.0|3468.0|3782.0|3772.0|3099.0|3117.0|4217.0|na|
 
-|Datasets|mnli_m|mnli_mm|ar|bg|de|el|en|es|fr|hi|ru|sw|th|tr|ur|vi|zh|
-| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
-|Accuracy|0.852|0.854|0.757|0.808|0.796|0.79|0.856|0.814|0.806|0.751|0.782|0.725|0.757|0.766|0.729|0.784|0.782|
-|Speed (text/sec)|2098.0|2170.0|3501.0|3324.0|3438.0|3174.0|3713.0|3500.0|3129.0|3042.0|3419.0|3468.0|3782.0|3772.0|3099.0|3117.0|4217.0|
+|Datasets|mnli_m|mnli_mm|
+| :---: | :---: | :---: | 
+|Accuracy|0.852|0.854|
+|Speed (text/sec)|2098.0|2170.0|
 
 
 ## Limitations and bias
 Please consult the original XLM-V paper and literature on different NLI datasets for potential biases. 
 
 ## Citation
-If you use this model, please cite: Laurer, Moritz, Wouter van Atteveldt, Andreu Salleras Casas, and Kasper Welbers. 2022. ‘Less Annotating, More Classifying – Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and BERT - NLI’. Preprint, June. Open Science Framework. https://osf.io/74b8k.
+If you use this model, please cite: Laurer, Moritz, Wouter van Atteveldt, Andreu Salleras Casas, and Kasper Welbers. 2022. 
+‘Less Annotating, More Classifying – Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and BERT - NLI’. 
+Preprint, June. Open Science Framework. https://osf.io/74b8k.
 
 ## Ideas for cooperation or questions?
 If you have questions or ideas for cooperation, contact me at m{dot}laurer{at}vu{dot}nl or [LinkedIn](https://www.linkedin.com/in/moritz-laurer/)