mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.1

This model is a fine-tuned version of asadfgglie/mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.0 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5335
• F1 Macro: 0.8675
• F1 Micro: 0.8692
• Accuracy Balanced: 0.8674
• Accuracy: 0.8692
• Precision Macro: 0.8677
• Recall Macro: 0.8674
• Precision Micro: 0.8692
• Recall Micro: 0.8692

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 128
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.06
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	F1 Macro	F1 Micro	Accuracy Balanced	Accuracy	Precision Macro	Recall Macro	Precision Micro	Recall Micro
0.1975	0.17	200	0.3474	0.8688	0.8708	0.8678	0.8708	0.8701	0.8678	0.8708	0.8708
0.1974	0.34	400	0.3580	0.8600	0.8624	0.8585	0.8624	0.8621	0.8585	0.8624	0.8624
0.2054	0.51	600	0.3616	0.8520	0.8565	0.8476	0.8565	0.8638	0.8476	0.8565	0.8565
0.2094	0.68	800	0.3772	0.8658	0.8687	0.8630	0.8687	0.8710	0.8630	0.8687	0.8687
0.2118	0.85	1000	0.3701	0.8729	0.8740	0.8747	0.8740	0.8719	0.8747	0.8740	0.8740
0.1948	1.02	1200	0.3778	0.8698	0.8714	0.8702	0.8714	0.8696	0.8702	0.8714	0.8714
0.1447	1.19	1400	0.3964	0.8666	0.8692	0.8642	0.8692	0.8706	0.8642	0.8692	0.8692
0.1723	1.35	1600	0.3855	0.8718	0.8735	0.8716	0.8735	0.8720	0.8716	0.8735	0.8735
0.1476	1.52	1800	0.4164	0.8637	0.8661	0.8620	0.8661	0.8661	0.8620	0.8661	0.8661
0.1515	1.69	2000	0.3958	0.8724	0.8740	0.8725	0.8740	0.8724	0.8725	0.8740	0.8740
0.1378	1.86	2200	0.4390	0.8694	0.8708	0.8699	0.8708	0.8689	0.8699	0.8708	0.8708
0.1332	2.03	2400	0.4535	0.8732	0.8745	0.8740	0.8745	0.8726	0.8740	0.8745	0.8745
0.0913	2.2	2600	0.5235	0.8638	0.8661	0.8625	0.8661	0.8656	0.8625	0.8661	0.8661
0.1076	2.37	2800	0.5339	0.8638	0.8661	0.8623	0.8661	0.8659	0.8623	0.8661	0.8661
0.09	2.54	3000	0.5388	0.8670	0.8687	0.8667	0.8687	0.8672	0.8667	0.8687	0.8687
0.0928	2.71	3200	0.5266	0.8649	0.8666	0.8648	0.8666	0.8650	0.8648	0.8666	0.8666
0.0805	2.88	3400	0.5433	0.8658	0.8677	0.8654	0.8677	0.8663	0.8654	0.8677	0.8677

Eval results

Datasets	asadfgglie/nli-zh-tw-all/test	asadfgglie/BanBan_2024-10-17-facial_expressions-nli	eval_dataset
Accuracy	0.87	0.955	0.869
Inference text/sec (RTX4060ti, batch=128)	36.0	256.0	32.0

Framework versions

• Transformers 4.33.3
• Pytorch 2.5.1+cu121
• Datasets 2.14.7
• Tokenizers 0.13.3