mDeBERTa-v3-base-xnli-multilingual-zeroshot-v4.0-only-nli-downsample

This model use same dataset with asadfgglie/mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.0, but training set was downsampled as 80% size of non-nli dataset asadfgglie/BanBan_2024-10-17-facial_expressions-nli.

This model is a fine-tuned version of MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4486
• F1 Macro: 0.8264
• F1 Micro: 0.8274
• Accuracy Balanced: 0.8270
• Accuracy: 0.8274
• Precision Macro: 0.8260
• Recall Macro: 0.8270
• Precision Micro: 0.8274
• Recall Micro: 0.8274

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: 2e-05
• train_batch_size: 16
• eval_batch_size: 128
• seed: 20241201
• 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
0.3242	1.69	200	0.4044	0.8308	0.8312	0.8322	0.8312	0.8306	0.8322	0.8312	0.8312

Eval results

Datasets	asadfgglie/nli-zh-tw-all/test	asadfgglie/BanBan_2024-10-17-facial_expressions-nli/test	eval_dataset	test_dataset
eval_loss	0.445	1.142	0.429	0.449
eval_f1_macro	0.827	0.505	0.83	0.826
eval_f1_micro	0.828	0.55	0.831	0.827
eval_accuracy_balanced	0.828	0.548	0.831	0.827
eval_accuracy	0.828	0.55	0.831	0.827
eval_precision_macro	0.827	0.575	0.83	0.826
eval_recall_macro	0.828	0.548	0.831	0.827
eval_precision_micro	0.828	0.55	0.831	0.827
eval_recall_micro	0.828	0.55	0.831	0.827
eval_runtime	275.581	4.734	54.573	209.065
eval_samples_per_second	30.844	199.853	31.151	32.526
eval_steps_per_second	0.243	1.69	0.257	0.258
epoch	2.99	2.99	2.99	2.99
Size of dataset	8500	946	1700	6800

Framework versions

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