QA-DeBERTa-MeanPooling-binary

This model is a fine-tuned version of microsoft/deberta-v3-large on the saiteki-kai/Beavertails-it dataset. It achieves the following results on the evaluation set:

• Loss: 0.3246
• Accuracy: 0.8645
• Unsafe Precision: 0.8883
• Unsafe Recall: 0.8653
• Unsafe F1: 0.8766
• Unsafe Fpr: 0.1366
• Unsafe Aucpr: 0.9567
• Safe Precision: 0.8363
• Safe Recall: 0.8634
• Safe F1: 0.8497
• Safe Fpr: 0.1347
• Safe Aucpr: 0.9241

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• 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: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Unsafe Precision	Unsafe Recall	Unsafe F1	Unsafe Fpr	Unsafe Aucpr	Safe Precision	Safe Recall	Safe F1	Safe Fpr	Safe Aucpr
0.3568	0.3001	5073	0.3409	0.8500	0.8498	0.8871	0.8681	0.1966	0.9480	0.8501	0.8034	0.8261	0.1129	0.9069
0.3161	0.6001	10146	0.3243	0.8592	0.8766	0.8693	0.8729	0.1536	0.9519	0.8377	0.8464	0.8420	0.1307	0.9135
0.304	0.9002	15219	0.3204	0.8609	0.8710	0.8805	0.8757	0.1636	0.9537	0.8480	0.8364	0.8422	0.1195	0.9169
0.3128	1.2002	20292	0.3311	0.8627	0.8932	0.8556	0.8740	0.1283	0.9532	0.8279	0.8717	0.8493	0.1444	0.9191
0.3198	1.5003	25365	0.3146	0.8624	0.8815	0.8695	0.8755	0.1466	0.9555	0.8391	0.8534	0.8462	0.1305	0.9225
0.2852	1.8003	30438	0.3174	0.8647	0.8937	0.8591	0.8761	0.1282	0.9568	0.8314	0.8718	0.8511	0.1409	0.9240
0.2694	2.1004	35511	0.3260	0.8642	0.8895	0.8633	0.8762	0.1346	0.9566	0.8346	0.8654	0.8497	0.1367	0.9231
0.3003	2.4004	40584	0.3265	0.8640	0.8850	0.8683	0.8766	0.1415	0.9568	0.8386	0.8585	0.8484	0.1317	0.9239
0.2723	2.7005	45657	0.3246	0.8644	0.8882	0.8653	0.8766	0.1366	0.9567	0.8363	0.8634	0.8496	0.1347	0.9241

Framework versions

• Transformers 4.57.1
• Pytorch 2.9.1+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1