victor-234/dhmsw-e-bert

DHMSW: E-BERT

Use

This model corresponds to E_BERT in our paper Targeting Transparency: Early Evidence on Mandatory Adoption of European Sustainability Reporting Standards. We use the model to classify sentences from firms' sustainability reports.

Base and metrics

This model is a fine-tuned version of distilbert/distilroberta-base on the esrs-sentences dataset. It achieves the following results on the evaluation set:

• Loss: 0.2439
• Accuracy: 0.9693
• F1: 0.9415
• Precision: 0.9430
• Recall: 0.9405

Training procedure

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: 2
• total_train_batch_size: 32
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 15
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Precision	Recall
0.5477	1.0	650	0.2079	0.9441	0.9009	0.9329	0.8768
0.1794	2.0	1300	0.1848	0.9485	0.9124	0.9029	0.9272
0.1159	3.0	1950	0.1852	0.9597	0.9148	0.9366	0.8953
0.0663	4.0	2600	0.1574	0.9658	0.9300	0.9279	0.9334
0.0442	5.0	3250	0.1778	0.9675	0.9400	0.9437	0.9390
0.0296	6.0	3900	0.2086	0.9662	0.9404	0.9410	0.9412
0.0151	7.0	4550	0.2398	0.9684	0.9428	0.9527	0.9358
0.0119	8.0	5200	0.2314	0.9706	0.9435	0.9443	0.9437
0.0133	9.0	5850	0.2175	0.9693	0.9384	0.9331	0.9442
0.0093	10.0	6500	0.2470	0.9667	0.9360	0.9323	0.9405
0.0075	11.0	7150	0.2206	0.9710	0.9462	0.9475	0.9454
0.0018	12.0	7800	0.2476	0.9684	0.9433	0.9517	0.9364
0.0031	13.0	8450	0.2496	0.9684	0.9398	0.9371	0.9429
0.0014	14.0	9100	0.2439	0.9693	0.9415	0.9430	0.9405

Framework versions

• Transformers 4.57.3
• Pytorch 2.8.0+cu128
• Datasets 4.4.2
• Tokenizers 0.22.2