Malay–English Sentiment Classifier (mBERT)
A three-class sentiment model developed for a 2025–2026 degree capstone on Malaysian COVID-19 social-media text. Labels: NEGATIVE (0), NEUTRAL (1), POSITIVE (2).
Current evaluated release
The corrected checkpoint is in corrected-2026-10, pinned at 9bc6645ce7a8e02761c59386ffd16969bed1bc55. The live app uses this exact release and verifies its weights checksum. The root-level model files are retained as the older checkpoint for rollback; loading the repository root alone does not load the corrected release.
Release protocol, preprocessing and limitations · Live app · Source and aggregate results
The corrected, shared split retains 6,539 unique inputs: 4,577 training, 981 validation and 981 test. Model selection used validation only. On the same internal test inputs, mBERT achieved 76.04% accuracy / 0.7557 macro-F1, compared with 66.46% / 0.6553 for TF-IDF logistic regression. Earlier development used this broader corpus; these are retrospective internal results, not independent external validation. These metrics apply only to the corrected checkpoint.
Correct usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
repo = "SiobhanGoh/CP2_Emotion_Model_New"
release = {"revision": "9bc6645ce7a8e02761c59386ffd16969bed1bc55",
"subfolder": "corrected-2026-10"}
tokenizer = AutoTokenizer.from_pretrained(repo, **release, fix_mistral_regex=False)
model = AutoModelForSequenceClassification.from_pretrained(repo, **release, use_safetensors=True)
model.eval()
Before inference, apply Unicode NFKC and casefolding, remove URLs and @mentions, remove hashtag markers and collapse whitespace. Reject empty cleaned inputs and truncate to 128 model tokens. The source app implements this preprocessing exactly as used during training.
Weights SHA-256: 771a19b8c14865f8d8d8c7adc7df95d859b08a866195f5b7bfe3ac44723ec61b.
Limits
This is an educational sentiment prototype. Sarcasm, slang, mixed languages, context and topics outside the training corpus can cause errors. Softmax scores are uncalibrated; they are not probabilities of correctness, emotion intensity, business risk or clinical assessments. YouTube comment summaries are not representative audience surveys. No raw social text is distributed here.
Author: Siobhan Goh Shze Mun.
- Downloads last month
- 23