hari1712/indicbert-sentiment

Fine-tuned IndicBERT sentiment classifier

This repository contains the fine-tuned IndicBERT sequence-classification checkpoint used by the IndicBERT text-classifier project. It predicts one of two labels, negative or positive, for Tamil, Hindi, and Malayalam text.

Model details

• Base checkpoint: ai4bharat/IndicBERTv2-MLM-only
• Task: binary sentiment classification
• Languages: Tamil, Hindi, Malayalam
• Classification head: mean pooling over active token representations followed by a small neural classifier
• Training implementation: the custom model class and training code in the linked project repository

The checkpoint was trained using the AI4Bharat IndicSentiment dataset. Review the upstream model and dataset terms, attribution, and licenses before redistribution or commercial use. This model card does not assign a license to the derived checkpoint.

Intended use and limitations

Use this model for exploratory positive/negative sentiment classification of short text in the listed languages. It is not a general-purpose language, moderation, or emotion-recognition model. Predictions may be unreliable for code-mixed text, other languages, domain-specific wording, sarcasm, or inputs unlike the training data. Evaluate it for your use case before relying on its outputs.

Use with this project

Set HF_MODEL_REPO_ID=hari1712/indicbert-sentiment in the environment where the project runs, or configure the same value as a Space variable. The project's model and tokenizer loading code is required because the classification head uses a custom architecture.