README.md

---
license: apache-2.0
tags:
- Cross-lingual-nlp
- zero-shot-transfer
- toxicity-analysis
- abuse-detection
- flag-user
- block-user
- multilinguality
- XLM-R
requirements:
- sentencepiece: (if not installed install using `pip install sentencepiece`)
---

# Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

This model aims to help developers, especially those with little to no experience in NLP, use our model directly to flag or block users from their platforms. Our model aims to work with any arbitrary language, as long as it is supported by the XLM-R vector space aligner embedder model. #Abuse detection #Toxicity analysis #Obscene language detection

Langauges supported:
- Afrikaans
- Albanian
- Amharic
- Arabic
- Armenian
- Assamese
- Azerbaijani
- Basque
- Belarusian
- Bengali
- Bhojpuri
- Bosnian
- Bulgarian
- Burmese
- Catalan
- Cebuano
- Chewa
- Chinese (Simplified)
- Chinese (Traditional)
- Chittagonian
- Corsican
- Croatian
- Czech
- Danish
- Deccan
- Dutch
- English
- Esperanto
- Estonian
- Filipino
- Finnish
- French
- Frisian
- Galician
- Georgian
- German
- Greek
- Gujarati
- Haitian Creole
- Hausa
- Hawaiian
- Hebrew
- Hindi
- Hmong
- Hungarian
- Icelandic
- Igbo
- Indonesian
- Irish
- Italian
- Japanese
- Javanese
- Kannada
- Kazakh
- Khmer
- Kinyarwanda
- Kirundi
- Korean
- Kurdish
- Kyrgyz
- Lao
- Latin
- Latvian
- Lithuanian
- Luxembourgish
- Macedonian
- Malagasy
- Malay
- Malayalam
- Maltese
- Maori
- Marathi
- Mongolian
- Nepali
- Norwegian
- Oriya
- Oromo
- Pashto
- Persian
- Polish
- Portuguese
- Punjabi
- Quechua
- Romanian
- Russian
- Samoan
- Scots Gaelic
- Serbian
- Shona
- Sindhi
- Sinhala
- Slovak
- Slovenian
- Somali
- Spanish
- Sundanese
- Swahili
- Swedish
- Tajik
- Tamil

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->



- **Developed by:** [Jayveersinh Raj](https://www.linkedin.com/in/jayveersinh-raj-67694222a/), [Khush Patel](https://www.linkedin.com/in/khush-patel-kp/)
- **Model type:** Cross-lingual-zero-shot-transfer
- **Language(s) (NLP):** Pytorch, ONNX
- **License:** apache-2.0

### Model Sources [optional]

<!-- Provide the basic links for the model. -->

- **Repository:** https://github.com/Jayveersinh-Raj/cross-lingual-zero-shot-transfer
- **Paper [optional]:** Everything is in the above github repository Make sure to give it a star if it is useful.
- **Demo [optional]:** 

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
This model aims to help developers, especially those with little to no experience in NLP, use our model directly to flag or block users from their platforms. Our model aims to work with any arbitrary language, as long as it is supported by the XLM-R vector space aligner embedder model. #Abuse detection #Toxicity analysis #Obscene language detection

### Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
Just use the model from hugging face directly


### Downstream Use [optional]

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
The model fine tuning is not needed the model already performs well. However can be fine tuned to add languages that are written with different scripts since our model does not perform on language with different script then the source. 


### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
This model does not work with a language written in different script. The transliteration layer has not been added yet.
Moreover, our model flags mostly severe toxicity, since toxicity is a subjective matter. However, in context of flagging or blocking users severty is very important, and our model is very well balanced in that aspect.


## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->
Toxicity is a subjective issue, however the model is very well balanced to flag mostly severe toxicity. The model has never flagged non toxic sentence as toxic. Its performance on non toxicity is 100%, making it a very good choice for the purpose of flagging or blocking users. In addition, if the language is very low resource, then the model might misclassify, but the performance is still good.


### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

## How to Get Started with the Model

Use the code below to get started with the model.


## Training Details

### Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
The training data involves the data from google jigsaw, and wikipidea. The training language is english, but zero shot mechanism is used to achieve multilinguality using vector space alignment.

### Training Procedure 

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

#### Preprocessing [optional]

We merged all the sub categories of toxicity to create a super category of toxicity, since all of them are severe, flaggable, and/or blockable.
Class imbalance was present, but state of the art transformer architecture can handle it well.

## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->
The model better than GPT4, and a human annotator that annotated the comments of test set as toxic. We arrived at this conclusion because 1. They were manually checked, and 2. On being GPT4 refused to generate toxic sentences, but on being passed the texts from test set where model flagged it non toxic, but were flagged toxic by user, GPT4 translated it, generated it, and said they were toxic, but they were not toxic enough to be blocked or flagged. Hence, our model is near to perfect in this regard. However, limitations, and risks should be taken into account.

### Testing Data, Factors & Metrics
1. Tested on human annotations 
2. Tested on GPT4 generated texts

#### Testing Data

<!-- This should link to a Data Card if possible. -->
The dataset is available on github

#### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->
Top-1 accuracy, since our data contains multiple langauges.

### Results
1. Tested on human annotations &rarr; 100% on non toxic sentences, better than human, as discussed in evaluation.
2. Tested on GPT4 generated texts &rarr; 100%

#### Summary
Our model is very good for the use case of flagging or blocking users with severe toxic comments, like using swear words or slangs. It is ideal for the purpose because it only flags severe toxicity, and 100% accurate on non-toxic comments. However all of the above should be taken into consideration before using it. It supports all the languages that are supported by XLM-R vector space aligner. The list is as follows: