This is a finetuned version of RoBERTa base, finetuned for detecting extreme hate (i.e extreme racism, homophobia, and suicide incitement)
This model is actively being improved and used in the AetherJr Discord bot.
Training Info
This model was trained on 121753 messages for 3 epochs using an NVIDIA RTX 3080 (10gb).
learning_rate: 3e-5
train_batch_size: 8
F1: 0.9423660
Recall: 0.9521694
Total Offensive/Bad: 1456
Total Normal: 120299
Dataset used
This model was trained on a collection of Discord messages obtained from various public chats across multiple servers, with the explicit permission of the server owners.
The dataset was chosen for its relevance to the model's objective of detecting extreme racism, homophobia, and suicide incitement in online communications. To ensure privacy, all data has been anonymized, with personal identifiers and links removed.
The dataset used for this model will not be publicly released.
Limitations and Bias
The dataset, primarily from English-speaking servers with intense hate content, may limit the model's ability to effectively recognize and understand non-English hate expressions.
Also, the hate data consists of mostly slurs, short 2-3 sentence hate and is highly directed at black/gay people, as such, it may not recognize more generic/subtle forms of hate.
Inference
The simplest way to use this model is using pipelines
⚠️ Make sure to pre-process your input text, removing links and ids as the model has a maximum input length of 512, you could also split the input into 512 long segments and do some math to calculate the score. It's up to you.
python
1import torch
2from transformers import pipeline
34device ='cuda:0'if torch.cuda.is_available()else'cpu'5classifier = pipeline(task="text-classification", model="imaether/roberta-offensive-junior", device=device)# add `top_k=None` if you want all scores to be returned67in_text ="this is a very cool and real example!"89model_output = classifier(in_text)10print(model_outputs)11# Output: [{'label': 'normal', 'score': 0.9999141693115234}]
