roberta-fnf-taglish-v1

RoBERTa-Tagalog fine-tuned for binary fake-news detection (real, fake).

Training setup

• Cluster-disjoint splits on cleaned FNF corpus
• Train-only Taglish paraphrase augmentation
• Base tokenizer: jcblaise/roberta-tagalog-base

Quickstart

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
repo = 'renshhhh/roberta-fnf-taglish-v1'
tok = AutoTokenizer.from_pretrained(repo, use_fast=True)
mdl = AutoModelForSequenceClassification.from_pretrained(repo)
text = 'Ito ay balitang halimbawa lang, hindi totoong artikulo.'
batch = tok(text, return_tensors='pt', truncation=True, max_length=256)
with torch.no_grad():
    probs = mdl(**batch).logits.softmax(-1).tolist()[0]
id2label = mdl.config.id2label
print({id2label[i]: float(probs[i]) for i in range(len(probs))})

Eval (cluster-disjoint test)

• Accuracy = 0.947
• Weighted F1 = 0.947