IndoBERT-lite sentiment classifier for the Free Nutritious Meals program (Makan Bergizi Gratis (MBG))

Classifies Indonesian tweets about the Free Nutritious Meals program (Makan Bergizi Gratis (MBG)) into negatif, netral or positif.

• Base model: indobenchmark/indobert-lite-base-p1 (IndoBERT-lite, ALBERT architecture, 11.7M parameters, MIT license)
• Fine-tuning data: 770 tweets (70% stratified split, seed 42) from tasyapndya/free_nutritious_meals_sentiment_x_id_2025
• Training: new 3-class head, AdamW (lr 2e-5, weight decay 0.01), batch 16, linear warmup 10%, up to 8 epochs, best epoch by validation F1-macro (epoch 5), fp16 on a T4 GPU
• Demo: https://nzlul03.github.io/mbg-sentiment-analysis/

Results (test set, 165 tweets)

McNemar test: this model vs Qwen3.5-0.8B zero-shot p = 0.0007; vs the TF-IDF baseline p = 0.327 (not significant on this small test set).

Files

Usage

Use BertTokenizer (this repo ships a WordPiece tokenizer.json). The original IndoBERT-lite card also loads the tokenizer with BertTokenizer; ALBERT's default SentencePiece tokenizer does not match this vocabulary.

Python:

from transformers import BertTokenizer, AutoModelForSequenceClassification, pipeline
repo = "Nzlul/indobert-mbg-sentiment"
clf = pipeline("text-classification", model=AutoModelForSequenceClassification.from_pretrained(repo),
               tokenizer=BertTokenizer.from_pretrained(repo), top_k=None)
clf("lagi lagi siswa keracunan menu mbg evaluasi dong")

JavaScript (Transformers.js, runs in the browser):

import { pipeline } from '@huggingface/transformers';
const clf = await pipeline('text-classification', 'Nzlul/indobert-mbg-sentiment', { dtype: 'fp32' });
await clf('lagi lagi siswa keracunan menu mbg evaluasi dong', { top_k: null });

Training text was lowercased with links, mentions and symbols removed; apply the same normalization for best results.

Limitations

Small dataset (1,100 tweets after cleaning) from a single platform and period (X, January–March 2025); labels are subjective, and the neutral class is the hardest (recall 0.68). Independent portfolio project, not affiliated with the Government of Indonesia or the National Nutrition Agency (Badan Gizi Nasional).