bert-base-japanese-jsnli

This model is a fine-tuned version of cl-tohoku/bert-base-japanese-v2 on the JSNLI dataset. It achieves the following results on the evaluation set:

• Loss: 0.2085
• Accuracy: 0.9288

How to use the model

Simple zero-shot classification pipeline

from transformers import pipeline

classifier = pipeline("zero-shot-classification", model="Formzu/bert-base-japanese-jsnli")

sequence_to_classify = "いつか世界を見る。"
candidate_labels = ['旅行', '料理', '踊り']
out = classifier(sequence_to_classify, candidate_labels, hypothesis_template="この例は{}です。")
print(out)
#{'sequence': 'いつか世界を見る。', 
# 'labels': ['旅行', '料理', '踊り'], 
# 'scores': [0.6758995652198792, 0.22110949456691742, 0.1029909998178482]}

NLI use-case

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model_name = "Formzu/bert-base-japanese-jsnli"
model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)

premise = "いつか世界を見る。"
label = '旅行'
hypothesis = f'この例は{label}です。'

input = tokenizer.encode(premise, hypothesis, return_tensors='pt').to(device)
with torch.no_grad():
    logits = model(input)["logits"][0]
    probs = logits.softmax(dim=-1)
    print(probs.cpu().numpy(), logits.cpu().numpy())
#[0.68940836 0.29482093 0.01577068] [ 1.7791482   0.92968255 -1.998533  ]

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.4054	1.0	16657	0.2141	0.9216
0.3297	2.0	33314	0.2145	0.9236
0.2645	3.0	49971	0.2085	0.9288

Framework versions

• Transformers 4.21.2
• Pytorch 1.12.1+cu116
• Datasets 2.4.0
• Tokenizers 0.12.1