README.md

metadata

base_model: asadfgglie/mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.0
tags:
  - generated_from_trainer
metrics:
  - accuracy
model-index:
  - name: mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.1-seed20241201
    results: []
datasets:
  - asadfgglie/nli-zh-tw-all
  - asadfgglie/BanBan_2024-10-17-facial_expressions-nli
language:
  - zh
pipeline_tag: zero-shot-classification

mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.1-seed20241201

This model use same hyper-parameter with asadfgglie/mDeBERTa-v3-base-xnli-multilingual-zeroshot-v1.1, except RANDOM_SEED.

Original version use RANDOM_SEED=42, this version use RANDOM_SEED=20241201.

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: 0.6134
• F1 Macro: 0.8616
• F1 Micro: 0.8634
• Accuracy Balanced: 0.8616
• Accuracy: 0.8634
• Precision Macro: 0.8616
• Recall Macro: 0.8616
• Precision Micro: 0.8634
• Recall Micro: 0.8634

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
0.2034	0.17	200	0.4241	0.8481	0.8518	0.8451	0.8518	0.8541	0.8451	0.8518	0.8518
0.219	0.34	400	0.4178	0.8608	0.8624	0.8615	0.8624	0.8602	0.8615	0.8624	0.8624
0.2142	0.51	600	0.3810	0.8572	0.8602	0.8548	0.8602	0.8613	0.8548	0.8602	0.8602
0.199	0.68	800	0.4314	0.8537	0.8571	0.8508	0.8571	0.8590	0.8508	0.8571	0.8571
0.2005	0.85	1000	0.4282	0.8572	0.8602	0.8547	0.8602	0.8615	0.8547	0.8602	0.8602
0.1846	1.02	1200	0.4631	0.8691	0.8703	0.8707	0.8703	0.8681	0.8707	0.8703	0.8703
0.154	1.19	1400	0.4922	0.8599	0.8613	0.8610	0.8613	0.8590	0.8610	0.8613	0.8613
0.1432	1.35	1600	0.5020	0.8540	0.8560	0.8540	0.8560	0.8541	0.8540	0.8560	0.8560
0.1335	1.52	1800	0.5313	0.8479	0.8507	0.8461	0.8507	0.8505	0.8461	0.8507	0.8507
0.1373	1.69	2000	0.5018	0.8546	0.8571	0.8533	0.8571	0.8563	0.8533	0.8571	0.8571
0.128	1.86	2200	0.4896	0.8644	0.8655	0.8665	0.8655	0.8633	0.8665	0.8655	0.8655
0.1257	2.03	2400	0.4922	0.8648	0.8666	0.8648	0.8666	0.8648	0.8648	0.8666	0.8666
0.0959	2.2	2600	0.5814	0.8589	0.8613	0.8576	0.8613	0.8606	0.8576	0.8613	0.8613
0.0918	2.37	2800	0.5987	0.8617	0.8634	0.8618	0.8634	0.8615	0.8618	0.8634	0.8634
0.0992	2.54	3000	0.6117	0.8631	0.8650	0.8629	0.8650	0.8634	0.8629	0.8650	0.8650
0.0897	2.71	3200	0.6191	0.8583	0.8602	0.8583	0.8602	0.8584	0.8583	0.8602	0.8602
0.1065	2.88	3400	0.6221	0.8625	0.8645	0.8619	0.8645	0.8631	0.8619	0.8645	0.8645

Eval result

Datasets	asadfgglie/nli-zh-tw-all/test	asadfgglie/BanBan_2024-10-17-facial_expressions-nli	eval_dataset
Accuracy	0.871	0.952	0.863
Inference text/sec (RTX4060ti, batch=128)	38.0	278.0	37.0

Framework versions

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