Views
No views yet
| 評価指標 | スコア |
|---|---|
| F値_NOT | 73.9 % |
| F値_GRY | 55.4 % |
| F値_OFF | 62.6 % |
| マクロ平均F値 | 64.0 % |
| 正解率 | 65.0 % |
| 正解ラベル \ 予測結果 | Not Offensive | Gray-area | Offensive |
|---|---|---|---|
| Not Offensive | 269 | 73 | 2 |
| Gray-area | 109 | 169 | 42 |
| Offensive | 6 | 48 | 82 |
1from transformers import AutoTokenizer, AutoModelForSequenceClassification
2import numpy as np
3
4tokenizer = AutoTokenizer.from_pretrained("studio-ousia/luke-japanese-large-lite")
5model = AutoModelForSequenceClassification.from_pretrained("TomokiFujihara/luke-japanese-large-lite-offensiveness-estimation", trust_remote_code=True)
6
7
8inputs = tokenizer.encode_plus(text, return_tensors='pt')
9outputs = model(inputs['input_ids'], inputs['attention_mask']).detach().numpy()[0][:3]
10
11minimum = np.min(outputs)
12if minimum < 0:
13 outputs = outputs - minimum
14score = outputs / np.sum(outputs)
15
16print(f'攻撃的でない発言: {score[0]:.1%},\nグレーゾーンの発言: {score[1]:.1%},\n攻撃的な発言: {score[2]:.1%}')
17