PhoBERT-HSD-Span

Fine-tuned from vinai/phobert-base on visolex/ViHOS for token-level hate/offensive span detection.

Model Details

• Base Model: vinai/phobert-base
• Dataset: visolex/ViHOS
• Fine-tuning: HuggingFace Transformers

Hyperparameters

• Batch size: 16
• Learning rate: 5e-5
• Epochs: 100
• Max sequence length: 128
• Early stopping: 5

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("visolex/phobert-hsd-span")
model = AutoModelForTokenClassification.from_pretrained("visolex/phobert-hsd-span")

text = "Nói cái lol . t thấy thô tục vl"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    outputs = model(**inputs)
logits = outputs.logits  # [batch, seq_len, num_labels]
# For binary: use sigmoid, for multi-class: use softmax+argmax
probs = torch.sigmoid(logits)
preds = (probs > 0.5).long().squeeze().tolist()  # [seq_len]
tokens = tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])

span_labels = [p[0] for p in preds]

span_tokens = [token for token, label in zip(tokens, span_labels) if label == 1 and token not in ['<s>', '</s>']]

print("Span tokens:", span_tokens)
print("Span text:", tokenizer.convert_tokens_to_string(span_tokens))