detoxify

repository·master·Indexed 23 days ago

A Python library for detecting toxic comments using PyTorch Lightning and Hugging Face Transformers. It provides pre-trained models for English and multilingual toxicity classification, including versions optimized to minimize unintended bias. Supported models include BERT, RoBERTa, XLM-RoBERTa, and ALBERT. The library features a Detoxify class for predictions, a CLI for command-line usage, and support for loading models via Torch Hub.

Tokens: 3K
Snippets: 11
Records: 19
Agent score: 79%

What's inside detoxify

Train the Toxic Comment Classification model
master
To train a model for the Toxic Comment Classification challenge, first combine the test data and labels using preprocessing_utils.py, then run the training script with the BERT configuration.
1. Combine test files:
python preprocessing_utils.py --test_csv <path_to_test_csv> --update_test 2. Start training: python train.py --config configs/Toxic_comment_classification_BERT.json
```
# combine test.csv and test_labels.csv
python preprocessing_utils.py --test_csv jigsaw_data/jigsaw-toxic-comment-classification-challenge/test.csv --update_test

python train.py --config configs/Toxic_comment_classification_BERT.json
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Evaluate the Unintended Bias in Toxicity model
master
To evaluate a model for the Unintended Bias challenge, use evaluate.py to get AUC scores, and then run the bias metric script to calculate the final specialized bias metric.
1. Run evaluation:
python evaluate.py --checkpoint <path_to_checkpoint> --test_csv <path_to_test_csv> 2. Compute final bias metric: python model_eval/compute_bias_metric.py
```
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv

# to get the final bias metric
python model_eval/compute_bias_metric.py
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Train the Unintended Bias in Toxicity model
master
To train a model specifically for the Unintended Bias in Toxicity challenge, use the RoBERTa combined configuration.
python train.py --config configs/Unintended_bias_toxic_comment_classification_RoBERTa_combined.json
```
python train.py --config configs/Unintended_bias_toxic_comment_classification_RoBERTa_combined.json
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Evaluate the Multilingual Toxic Comment Classification model
master
Evaluate a trained checkpoint for the Multilingual challenge using the evaluate.py script. This challenge is evaluated on the AUC score of the main toxic label.
python evaluate.py --checkpoint <path_to_checkpoint> --test_csv <path_to_test_csv>
```
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Evaluate the Toxic Comment Classification model
master
Evaluate a trained checkpoint for the Toxic Comment Classification challenge using the evaluate.py script. This challenge is evaluated on the mean AUC score of all labels.
python evaluate.py --checkpoint <path_to_checkpoint> --test_csv <path_to_test_csv>
```
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
https://github.com/unitaryai/detoxify/blob/master/README.md

Perform quick toxicity predictions with Detoxify

master

Use the Detoxify class to load a model and predict toxicity scores for a single string or a list of strings.

Supported model names include:

original (BERT-based)
unbiased (RoBERTa-based)
multilingual (XLM-RoBERTa-based)
original-small (Albert-based)
unbiased-small (Albert-based)

You can also specify a device (e.g., 'cuda') to allocate the model to a specific hardware device.

from detoxify import Detoxify

# Predict on a single string
results = Detoxify('original').predict('example text')

# Predict on a list of strings
results = Detoxify('unbiased').predict(['example text 1','example text 2'])

# Use the multilingual model
results = Detoxify('multilingual').predict(['example text','exemple de texte'])

# Specify a device (e.g., GPU)
model = Detoxify('original', device='cuda')

from detoxify import Detoxify

# each model takes in either a string or a list of strings

results = Detoxify('original').predict('example text')

results = Detoxify('unbiased').predict(['example text 1','example text 2'])

results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])

# to specify the device the model will be allocated on (defaults to cpu), accepts any torch.device input

model = Detoxify('original', device='cuda')

# optional to display results nicely (will need to pip install pandas)

import pandas as pd

print(pd.DataFrame(results, index=input_text).round(5))

Install detoxify via pip
master
You can install the detoxify library directly from PyPI using pip.
```
pip install detoxify
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Train the Multilingual Toxic Comment Classification model
master
To train the multilingual model (supporting French, Spanish, Italian, Portuguese, Turkish, and Russian), first preprocess the multilingual test data and then run the training script with the XLMR configuration.
1. Combine test files:
python preprocessing_utils.py --test_csv <path_to_multilingual_test_csv> --update_test 2. Start training: python train.py --config configs/Multilingual_toxic_comment_classification_XLMR.json
```
# combine test.csv and test_labels.csv
python preprocessing_utils.py --test_csv jigsaw_data/jigsaw-multilingual-toxic-comment-classification/test.csv --update_test

python train.py --config configs/Multilingual_toxic_comment_classification_XLMR.json
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Monitor training progress with TensorBoard
master
You can monitor the training progress by running TensorBoard pointing to the ./saved directory where logs are stored.
tensorboard --logdir=./saved
```
tensorboard --logdir=./saved
```
https://github.com/unitaryai/detoxify/blob/master/README.md
Understand Detoxify class names and standardization
master
Detoxify standardizes class names across its different models to ensure consistent API usage. When you receive results from predict(), the keys in the returned dictionary will follow these mappings:
Original Name Standardized Name
toxic toxicity
identity_hate identity_attack
severe_toxic severe_toxicity
https://github.com/unitaryai/detoxify/blob/master/detoxify/detoxify.py
Load Detoxify models via Torch Hub
master
You can load pre-trained Detoxify checkpoints directly using torch.hub.load with the following names:
- toxic_bert (for the original model)
- unbiased_toxic_roberta (for the unbiased model)
- multilingual_toxic_xlm_r (for the multilingual model)
```
import torch
model = torch.hub.load('unitaryai/detoxify','toxic_bert')
```
```
model = torch.hub.load('unitaryai/detoxify','toxic_bert')
```
https://github.com/unitaryai/detoxify/blob/master/README.md

Original Name	Standardized Name
`toxic`	`toxicity`
`identity_hate`	`identity_attack`
`severe_toxic`	`severe_toxicity`

Reference: Detoxify model names and types

master

Model name	Transformer type	Data from
`original`	`bert-base-uncased`	Toxic Comment Classification Challenge
`unbiased`	`roberta-base`	Unintended Bias in Toxicity Classification
`multilingual`	`xlm-roberta-base`	Multilingual Toxic Comment Classification