Adding model cards for 5 models (#6703)

* Added model cards for 4 models

Added model cards for:
- roberta-base-bulgarian
- roberta-base-bulgarian-pos
- roberta-small-bulgarian
- roberta-small-bulgarian-pos

* fixed link text

* Update README.md

* Create README.md

* removed trailing bracket

* Add language metadata
Co-authored-by: Julien Chaumond <chaumond@gmail.com>

model_cards/iarfmoose/bert-base-cased-qa-evaluator/README.md

+# BERT-base-cased-qa-evaluator
+
+This model takes a question answer pair as an input and outputs a value representing its prediction about whether the input was a valid question and answer pair or not. The model is a pretrained [BERT-base-cased](https://huggingface.co/bert-base-cased) with a sequence classification head.
+
+## Intended uses
+
+The QA evaluator was originally designed to be used with the [t5-base-question-generator](https://huggingface.co/iarfmoose/t5-base-question-generator) for evaluating the quality of generated questions. 
+
+The input for the QA evaluator follows the format for `BertForSequenceClassification`, but using the question and answer as the two sequences. Inputs should take the following format:
+```
+[CLS] <question> [SEP] <answer [SEP]
+```
+
+## Limitations and bias
+
+The model is trained to evaluate if a question and answer are semantically related, but cannot determine whether an answer is actually true/correct or not.
+
+## Training data
+
+The training data was made up of question-answer pairs from the following datasets: 
+- [SQuAD](https://rajpurkar.github.io/SQuAD-explorer/)
+- [RACE](http://www.cs.cmu.edu/~glai1/data/race/)
+- [CoQA](https://stanfordnlp.github.io/coqa/)
+- [MSMARCO](https://microsoft.github.io/msmarco/)
+
+## Training procedure
+
+The question and answer were concatenated 50% of the time. In the other 50% of the time a corruption operation was performed (either swapping the answer for an unrelated answer, or by copying part of the question into the answer). The model was then trained to predict whether the input sequence represented one of the original QA pairs or a corrupted input.

model_cards/iarfmoose/roberta-base-bulgarian-pos/README.md

+---
+language: bg
+---
+
+# RoBERTa-base-bulgarian-POS
+
+
+The RoBERTa model was originally introduced in [this paper](https://arxiv.org/abs/1907.11692). This model is a version of [RoBERTa-base-Bulgarian](https://huggingface.co/iarfmoose/roberta-base-bulgarian) fine-tuned for part-of-speech tagging.
+
+## Intended uses
+
+The model can be used to predict part-of-speech tags in Bulgarian text. Since the tokenizer uses byte-pair encoding, each word in the text may be split into more than one token. When predicting POS-tags, the last token from each word can be used. Using the last token was found to slightly outperform predictions based on the first token.
+
+An example of this can be found [here](https://github.com/iarfmoose/bulgarian-nlp/blob/master/models/postagger.py).
+
+## Limitations and bias
+
+The pretraining data is unfiltered text from the internet and may contain all sorts of biases.
+
+## Training data
+
+In addition to the pretraining data used in [RoBERTa-base-Bulgarian]([RoBERTa-base-Bulgarian](https://huggingface.co/iarfmoose/roberta-base-bulgarian)), the model was trained on the UPOS tags from [UD_Bulgarian-BTB](https://github.com/UniversalDependencies/UD_Bulgarian-BTB).
+
+## Training procedure
+
+The model was trained for 5 epochs over the training set. The loss was calculated based on label predictions for the last POS-tag for each word. The model achieves 97% on the test set.

model_cards/iarfmoose/roberta-base-bulgarian/README.md

+---
+language: bg
+---
+
+# RoBERTa-base-bulgarian
+
+
+The RoBERTa model was originally introduced in [this paper](https://arxiv.org/abs/1907.11692). This is a version of [RoBERTa-base](https://huggingface.co/roberta-base) pretrained on Bulgarian text.
+
+## Intended uses
+
+This model can be used for cloze tasks (masked language modeling) or finetuned on other tasks in Bulgarian.
+
+## Limitations and bias
+
+The training data is unfiltered text from the internet and may contain all sorts of biases.
+
+## Training data
+
+This model was trained on the following data:
+- [bg_dedup from OSCAR](https://oscar-corpus.com/)
+- [Newscrawl 1 million sentences 2017 from Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download/bulgarian)
+- [Wikipedia 1 million sentences 2016 from Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download/bulgarian)
+
+## Training procedure
+
+The model was pretrained using a masked language-modeling objective with dynamic masking as described [here](https://huggingface.co/roberta-base#preprocessing)
+
+It was trained for 200k steps. The batch size was limited to 8 due to GPU memory limitations.

model_cards/iarfmoose/roberta-small-bulgarian-pos/README.md

+---
+language: bg
+---
+
+# RoBERTa-small-bulgarian-POS
+
+
+The RoBERTa model was originally introduced in [this paper](https://arxiv.org/abs/1907.11692). This model is a version of [RoBERTa-small-Bulgarian](https://huggingface.co/iarfmoose/roberta-small-bulgarian) fine-tuned for part-of-speech tagging.
+
+## Intended uses
+
+The model can be used to predict part-of-speech tags in Bulgarian text. Since the tokenizer uses byte-pair encoding, each word in the text may be split into more than one token. When predicting POS-tags, the last token from each word can be used. Using the last token was found to slightly outperform predictions based on the first token.
+
+An example of this can be found [here](https://github.com/iarfmoose/bulgarian-nlp/blob/master/models/postagger.py).
+
+## Limitations and bias
+
+The pretraining data is unfiltered text from the internet and may contain all sorts of biases.
+
+## Training data
+
+In addition to the pretraining data used in [RoBERTa-base-Bulgarian]([RoBERTa-base-Bulgarian](https://huggingface.co/iarfmoose/roberta-base-bulgarian)), the model was trained on the UPOS tags from (UD_Bulgarian-BTB)[https://github.com/UniversalDependencies/UD_Bulgarian-BTB].
+
+## Training procedure
+
+The model was trained for 5 epochs over the training set. The loss was calculated based on label predictions for the last POS-tag for each word. The model achieves 98% on the test set.

model_cards/iarfmoose/roberta-small-bulgarian/README.md

+---
+language: bg
+---
+
+# RoBERTa-small-bulgarian
+
+
+The RoBERTa model was originally introduced in [this paper](https://arxiv.org/abs/1907.11692). This is a smaller version of [RoBERTa-base-bulgarian](https://huggingface.co/iarfmoose/roberta-small-bulgarian) with only 6 hidden layers, but similar performance.
+
+## Intended uses
+
+This model can be used for cloze tasks (masked language modeling) or finetuned on other tasks in Bulgarian.
+
+## Limitations and bias
+
+The training data is unfiltered text from the internet and may contain all sorts of biases.
+
+## Training data
+
+This model was trained on the following data:
+- [bg_dedup from OSCAR](https://oscar-corpus.com/)
+- [Newscrawl 1 million sentences 2017 from Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download/bulgarian)
+- [Wikipedia 1 million sentences 2016 from Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download/bulgarian)
+
+## Training procedure
+
+The model was pretrained using a masked language-modeling objective with dynamic masking as described [here](https://huggingface.co/roberta-base#preprocessing)
+
+It was trained for 160k steps. The batch size was limited to 8 due to GPU memory limitations.