ymoslem/ModernBERT-base-qe-v1

Quality Estimation for Machine Translation

This model is a fine-tuned version of answerdotai/ModernBERT-base on the ymoslem/tokenized-wmt-da-human-evaluation dataset. It achieves the following results on the evaluation set:

• Loss: 0.0571

Model description

This model is for reference-free, sentence level quality estimation (QE) of machine translation (MT) systems. The long-context / document-level model can be found at: ModernBERT-base-long-context-qe-v1, which is trained on a long-context / document-level QE dataset ymoslem/wmt-da-human-evaluation-long-context

Training and evaluation data

This model is trained on the sentence-level quality estimation dataset: ymoslem/wmt-da-human-evaluation

Training procedure

This version of the model uses the full lengthtokenizer.model_max_length=8192, but it is still trained on a sentence-level QE dataset ymoslem/wmt-da-human-evaluation

The long-context / document-level model can be found at: ModernBERT-base-long-context-qe-v1, which is trained on a long-context / document-level QE dataset ymoslem/wmt-da-human-evaluation-long-context

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 128
• eval_batch_size: 128
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss
0.0686	0.1004	1000	0.0712
0.0652	0.2007	2000	0.0687
0.0648	0.3011	3000	0.0623
0.0609	0.4015	4000	0.0600
0.0585	0.5019	5000	0.0603
0.0588	0.6022	6000	0.0589
0.0592	0.7026	7000	0.0581
0.0585	0.8030	8000	0.0574
0.0588	0.9033	9000	0.0572
0.0563	1.0037	10000	0.0571

Framework versions

• Transformers 4.48.1
• Pytorch 2.4.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0