gulamsakaria/commentlens-banglishbert

CommentLens — Bangla/Banglish 5-Class Comment Classifier

Fine-tuned csebuetnlp/banglishbert (ELECTRA) for classifying Bangla/Banglish social media comments into 5 categories, with class-weighted loss to handle severe label imbalance.

Classes: claim | general | opinion | spam-scam | toxic

Training data

• 32,632 labeled Bangla/Banglish comments (CommentLens dataset)
• Class distribution: general 77% / opinion 18% / toxic 4% / spam-scam 0.6% / claim 0.6%
• 6 epochs, batch size 16, max length 128, LR 2e-5, class-weighted cross-entropy loss

Validation results

Class	Precision	Recall	F1	Support
claim	0.52	0.56	0.54	27
general	0.97	0.95	0.96	3768
opinion	0.84	0.92	0.88	875
spam-scam	0.96	0.93	0.95	28
toxic	0.77	0.79	0.78	197

Overall accuracy: 93% (macro avg F1: 0.82, weighted avg F1: 0.93) on a 4,895-row validation set.

Rare classes (claim, spam-scam) have low support and should be treated with caution in production.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("gulamsakaria/commentlens-banglishbert")
model = AutoModelForSequenceClassification.from_pretrained("gulamsakaria/commentlens-banglishbert")

text = "এই পেজে সবাই মিথ্যা কথা বলে"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)
pred = model.config.id2label[int(probs.argmax())]
print(pred, float(probs.max()))

Part of CommentLens

Built by Gulam Sakaria as part of the CommentLens project on Bangla comment moderation and misinformation detection.