eliasalbouzidi/distilroberta-nsfw-text-classifier

Model Card

This model is designed to categorize text into two classes: "safe", or "nsfw" (not safe for work), which makes it suitable for content moderation and filtering applications.

The model was trained using a dataset containing 190,000 labeled text samples, distributed among the two classes of "safe" and "nsfw".

The model is based on the Distilroberta-base model.

In terms of performance, the model has achieved a score of 0.975 for F1 (40K exemples).

To improve the performance of the model, it is necessary to preprocess the input text. You can refer to the preprocess function in the app.py file in the following space: https://huggingface.co/spaces/eliasalbouzidi/distilbert-nsfw-text-classifier.

Model Description

The model can be used directly to classify text into one of the two classes. It takes in a string of text as input and outputs a probability distribution over the two classes. The class with the highest probability is selected as the predicted class.

• Developed by: Centrale Supélec Students
• Model type: 82M
• Language(s) (NLP): English
• License: apache-2.0

Uses

The model can be integrated into larger systems for content moderation or filtering.

Training Data

The training data for finetuning the text classification model consists of a large corpus of text labeled with one of the two classes: "safe" and "nsfw". The dataset contains a total of 190,000 examples, which are distributed as follows:

117,000 examples labeled as "safe"

63,000 examples labeled as "nsfw"

It was assembled by scraping data from the web and utilizing existing open-source datasets. A significant portion of the dataset consists of descriptions for images and scenes. The primary objective was to prevent diffusers from generating NSFW content but it can be used for other moderation purposes.

You can access the dataset : https://huggingface.co/datasets/eliasalbouzidi/NSFW-Safe-Dataset

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 600
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Fbeta 1.6	False positive rate	False negative rate	Precision	Recall
0.3325	0.0998	586	0.1839	0.9479	0.9280	0.9132	0.0202	0.1052	0.9638	0.8948
0.1161	0.1997	1172	0.0958	0.9682	0.9576	0.9577	0.0256	0.0422	0.9574	0.9578
0.1006	0.2995	1758	0.0959	0.9716	0.9616	0.9554	0.0140	0.0524	0.9761	0.9476
0.0984	0.3994	2344	0.0776	0.9717	0.9628	0.9674	0.0293	0.0265	0.9523	0.9735
0.0897	0.4992	2930	0.0994	0.9676	0.9580	0.9696	0.0428	0.0151	0.9326	0.9849
0.0856	0.5991	3516	0.0889	0.9751	0.9670	0.9684	0.0219	0.0299	0.9638	0.9701
0.0779	0.6989	4102	0.0842	0.9762	0.9681	0.9652	0.0149	0.0385	0.9749	0.9615
0.0821	0.7988	4688	0.0702	0.9768	0.9693	0.9723	0.0228	0.0239	0.9626	0.9761
0.073	0.8986	5274	0.0773	0.9776	0.9703	0.9727	0.0213	0.0243	0.9650	0.9757
0.0775	0.9985	5860	0.0709	0.9774	0.9701	0.9721	0.0210	0.0252	0.9653	0.9748
0.0627	1.0983	6446	0.0646	0.9789	0.9720	0.9737	0.0193	0.0240	0.9681	0.9760
0.0648	1.1982	7032	0.0729	0.9787	0.9716	0.9708	0.0158	0.0303	0.9735	0.9697
0.0592	1.2980	7618	0.0733	0.9802	0.9735	0.9725	0.0144	0.0289	0.9760	0.9711
0.0585	1.3979	8204	0.0764	0.9790	0.9720	0.9723	0.0173	0.0273	0.9713	0.9727
0.0579	1.4977	8790	0.0691	0.9782	0.9712	0.9739	0.0214	0.0225	0.9649	0.9775
0.0584	1.5975	9376	0.0739	0.9797	0.9732	0.9755	0.0195	0.0215	0.9679	0.9785
0.0564	1.6974	9962	0.0749	0.9774	0.9703	0.9757	0.0257	0.0173	0.9583	0.9827
0.0582	1.7972	10548	0.0721	0.9804	0.9739	0.9742	0.0162	0.0253	0.9730	0.9747
0.0576	1.8971	11134	0.0746	0.9799	0.9732	0.9734	0.0163	0.0264	0.9729	0.9736
0.0546	1.9969	11720	0.0758	0.9804	0.9739	0.9736	0.0152	0.0269	0.9747	0.9731
0.0431	2.0968	12306	0.0755	0.9805	0.9741	0.9762	0.0186	0.0211	0.9694	0.9789
0.0464	2.1966	12892	0.0785	0.9802	0.9737	0.9735	0.0156	0.0266	0.9740	0.9734
0.0443	2.2965	13478	0.0763	0.9811	0.9748	0.9734	0.0132	0.0283	0.9779	0.9717
0.0426	2.3963	14064	0.0753	0.9812	0.9750	0.9760	0.0165	0.0227	0.9727	0.9773
0.0413	2.4962	14650	0.0750	0.9811	0.9748	0.9760	0.0168	0.0225	0.9722	0.9775
0.0442	2.5960	15236	0.0756	0.9813	0.9752	0.9766	0.0169	0.0216	0.9720	0.9784
0.043	2.6959	15822	0.0810	0.9814	0.9750	0.9729	0.0119	0.0299	0.98	0.9701
0.0433	2.7957	16408	0.0783	0.9814	0.9751	0.9733	0.0125	0.0289	0.9790	0.9711
0.0398	2.8956	16994	0.0736	0.9814	0.9752	0.9766	0.0169	0.0216	0.9721	0.9784
0.0431	2.9954	17580	0.0757	0.9816	0.9754	0.9757	0.0151	0.0240	0.9749	0.9760
We selected the checkpoint with the highest F-beta1.6 score.

Framework versions

• Transformers 4.40.1
• Pytorch 2.3.0+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1

Out-of-Scope Use

It should not be used for any illegal activities.

Bias, Risks, and Limitations

The model may exhibit biases based on the training data used. It may not perform well on text that is written in languages other than English. It may also struggle with sarcasm, irony, or other forms of figurative language. The model may produce false positives or false negatives, which could lead to incorrect categorization of text.

Recommendations

Users should be aware of the limitations and biases of the model and use it accordingly. They should also be prepared to handle false positives and false negatives. It is recommended to fine-tune the model for specific downstream tasks and to evaluate its performance on relevant datasets.

Load model directly

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("eliasalbouzidi/distilroberta-nsfw-text-classifier")

model = AutoModelForSequenceClassification.from_pretrained("eliasalbouzidi/distilroberta-nsfw-text-classifier")

Use a pipeline

from transformers import pipeline

pipe = pipeline("text-classification", model="eliasalbouzidi/distilroberta-nsfw-text-classifier")

Contact

Please reach out to eliasalbouzidi@gmail.com if you have any questions or feedback.