metadata
language:
- en
tags:
- text-classification
metrics:
- accuracy
datasets:
- snli-1.0
- multi-nli-1.0
- nli-fever
- anli-v1.0
deberta-v3-large-snli_mnli_fever_anli_R1_R2_R3-nli
Datasets
Based on microsoft/deberta-v3-large, this model was trained on the snli-v1.0, multi-nli-1.0, nli-fever and anli-1.0-r1/anli-1.0-r2/anli-1.0-r3 datasets, with the training weights of 1,1,1,10,20,10 respectively.
The training codes are mostly referenced from: https://github.com/facebookresearch/anli
Hyperparameters
learning_rate: 1e-5
max_length: 156
batch_size: 16
warmup_ratio: 0.1
weight_decay: 0.0
num_epochs: 2
dev_results
| snli-v1.0 | multi-nli-1.0-m | multi-nli-1.0-mm | anli-1.0-r1 | anli-1.0-r2 | anli-1.0-r3 |
|---|---|---|---|---|---|
| 0.938 | 0.914 | 0.912 | 0.796 | 0.627 | 0.610 |
test_results
Results of the test sets are shown together with some other official pre-trained model checkpoints.
| Model | snli-v1.0 | anli-1.0-r1 | anli-1.0-r2 | anli-1.0-r3 |
|---|---|---|---|---|
| ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli | - | 0.736 | 0.493 | 0.455 |
| ynie/xlnet-large-cased-snli_mnli_fever_anli_R1_R2_R3-nli | - | 0.700 | 0.514 | 0.498 |
| ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nli | - | 0.736 | 0.586 | 0.534 |
| deberta-v3-large-snli_mnli_fever_anli_R1_R2_R3-nli | 0.929 | 0.775 | 0.636 | 0.612 |