edu-modernbert

This model is a fine-tuned version of answerdotai/ModernBERT-base on the HuggingFaceFW/fineweb-edu-llama3-annotations dataset. It achieves the following results on the evaluation set:

• Loss: 0.2453
• Precision: 0.5901
• Recall: 0.5245
• F1: 0.5504
• Accuracy: 0.7508
• Binary Precision: 0.8168
• Binary Recall: 0.6856
• Binary F1: 0.7455
• Binary Accuracy: 0.9578

Note: the binary classification score is calculated by thresholding at 3 i.e (0-2 -> 0, 3-5 -> 1).

In comparison the reproduced version of HuggingFaceFW/fineweb-edu-classifier achieves:

• Loss: 0.2475
• Precision: 0.5595
• Recall: 0.4360
• F1: 0.4704
• Accuracy: 0.7123
• Binary Precision: 0.7781
• Binary Recall: 0.5566
• Binary F1: 0.6490
• Binary Accuracy: 0.9457

Note: one difference is that ModernBERT-base is fully trained while the original classifier trains only the regression head..

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 256
• eval_batch_size: 256
• seed: 0
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20(totally not needed, 3 epochs already achieve great results)

Framework versions

• Transformers 4.48.0.dev0
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0