Open models
Collection
Finetuned open source models.
•
3 items
•
Updated
Safe for Work Classifier Model for Malaysian Data
Current version supports Malay. We are working towards supporting malay, english and indo.
Base Model finetuned from https://huggingface.co/mesolitica/malaysian-mistral-191M-MLM-512 with Malaysian NSFW data.
Data Source: https://huggingface.co/datasets/malaysia-ai/Malaysian-NSFW
Github Repo: https://github.com/malaysia-ai/sfw-classifier
Project Board: https://github.com/orgs/malaysia-ai/projects/6
Current Labels Available:
from classifier import MistralForSequenceClassification
from transformers import AutoTokenizer
from transformers import pipeline
model = MistralForSequenceClassification.from_pretrained('malaysia-ai/malaysian-sfw-classifier')
tokenizer = AutoTokenizer.from_pretrained('malaysia-ai/malaysian-sfw-classifier')
pipe = pipeline("text-classification",
tokenizer = tokenizer,
model=model)
input_str = ["INSERT_INPUT_0", "INSERT_INPUT_1"]
print(pipe(input_str))
precision recall f1-score support
racist 0.87619 0.91390 0.89465 1719
religion insult 0.88533 0.85813 0.87152 3320
psychiatric or mental illness 0.94224 0.87020 0.90479 5624
sexist 0.77146 0.82234 0.79609 1486
harassment 0.81935 0.87460 0.84608 949
porn 0.95047 0.97546 0.96280 1141
safe for work 0.83471 0.90741 0.86954 3456
self-harm 0.81796 0.95906 0.88291 342
violence 0.84317 0.78786 0.81457 1433
accuracy 0.87684 19470
macro avg 0.86010 0.88544 0.87144 19470
weighted avg 0.87960 0.87684 0.87718 19470
@misc{razak2024adaptingsafeforworkclassifiermalaysian,
title={Adapting Safe-for-Work Classifier for Malaysian Language Text: Enhancing Alignment in LLM-Ops Framework},
author={Aisyah Razak and Ariff Nazhan and Kamarul Adha and Wan Adzhar Faiq Adzlan and Mas Aisyah Ahmad and Ammar Azman},
year={2024},
eprint={2407.20729},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.20729},
}