microsoft-swinv2-small-patch4-window16-256-finetuned-xblockm

This model is a fine-tuned version of microsoft/swinv2-small-patch4-window16-256 on the howdyaendra/xblock-social-screenshots dataset. It achieves the following results on the evaluation set:

• Loss: 0.1252
• Roc Auc: 0.9535

Model description

This model is trained on several thousand screenshots reported to the XBlock 3rd-party Bluesky labeller service. It is intended to be used to label Bluesky posts that have screenshots from social media sites embedded in them. Please also see aendra-rininsland/xblock.

Intended uses & limitations

Screenshot moderation

Training and evaluation data

20% split of 1618 images

Training procedure

See notebook.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 8

Training results

Training Loss	Epoch	Step	Validation Loss	Roc Auc
0.4357	0.9877	20	0.2544	0.7784
0.2027	1.9753	40	0.2016	0.8431
0.1743	2.9630	60	0.1701	0.8912
0.1625	4.0	81	0.1677	0.9083
0.1321	4.9877	101	0.1447	0.9246
0.1155	5.9753	121	0.1418	0.9311
0.0959	6.9630	141	0.1381	0.9460
0.0788	7.9012	160	0.1252	0.9535

Framework versions

• Transformers 4.44.1
• Pytorch 2.2.2
• Datasets 3.0.1
• Tokenizers 0.19.1