README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - toxic
  - toxicity
  - offensive language
  - hate speech

Text Classification GoEmotions

This model is a fined-tuned version of MiniLMv2-L6-H384 on the on the Jigsaw 1st Kaggle competition dataset using unitary/toxic-bert as teacher model. The quantized version in ONNX format can be found here.

The model with two labels only (toxicity and severe toxicity) is here

Load the Model

from transformers import pipeline

pipe = pipeline(model='minuva/MiniLMv2-toxic-jigsaw', task='text-classification')
pipe("This is pure trash")
# [{'label': 'toxic', 'score': 0.9383478164672852}]

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-05
• train_batch_size: 48
• eval_batch_size: 48
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• warmup_ratio: 0.1

Metrics (comparison with teacher model)

Teacher (params)	Student (params)	Set (metric)	Score (teacher)	Score (student)
unitary/toxic-bert (110M)	MiniLMv2-toxic-jigsaw (23M)	Test (ROC_AUC)	0.98636	0.98600

Deployment

Check out fast-nlp-text-toxicity repository for a FastAPI based server to deploy this model in CPU devices.