KordAI/III-BERT-Wisesight-Sentiment

Wisesight Thai sentiment — III-BERT

Fine-tuned downstream checkpoint built from KordAI/III-BERT.

• Task: Wisesight Thai sentiment
• Labels (4): pos, neu, neg, q
• Base model: KordAI/III-BERT
• Custom code: this repository contains III-BERT model code. Load it with trust_remote_code=True.

Test metrics

Metric	Value
test_loss	0.729800
test_accuracy	0.725200
test_f1_macro	0.606600
test_f1_weighted	0.717400
test_f1_neg	0.749500
test_f1_neu	0.781200
test_f1_pos	0.518000
test_f1_q	0.377800

The Wisesight scores are the held-out test results from the local training run.

Loading

from transformers import CamembertTokenizer, AutoModelForSequenceClassification

model_id = "KordAI/III-BERT-Wisesight-Sentiment"
tokenizer = CamembertTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(
    model_id,
    trust_remote_code=True,
)

For token classification, the model outputs one label distribution per tokenized subword. Map subword predictions back to words according to your tokenizer alignment strategy.