scandi-fine-web-cleaner

This model is a demo classifier for identifying problematic content (incorrect language, garbled text) in Danish and Swedish web text. It was created as part of a blog post exploring how to filter web data using community annotations. The model was created by fine-tuning FacebookAI/xlm-roberta-base on the data-is-better-together/fineweb-c dataset.

It achieves the following results on the evaluation set:

• Precision: 0.9524 (95.2%)
• Recall: 0.7018 (70.2%)
• F1: 0.8081
• AUC-ROC: 0.9648

Intended uses & limitations

The model is intended to be used as a preliminary filter for web text to help improve annotation efficiency. It has only been tested on Danish and Swedish content. The high precision (95.2%) means false positives are rare, while the recall (70.2%) indicates it catches most problematic content.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Auc Roc	Balanced Accuracy	Average Precision
0.3165	1.0	100	0.2333	0.95	0.6667	0.7835	0.8099	0.8304	0.7721
0.1929	2.0	200	0.1359	0.9130	0.7368	0.8155	0.9778	0.8626	0.9105
0.1775	3.0	300	0.2245	0.9268	0.6667	0.7755	0.9481	0.8290	0.8721
0.1553	4.0	400	0.1816	0.9524	0.7018	0.8081	0.9648	0.8480	0.8906

Framework versions

• Transformers 4.48.0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0