MovieCert Marathi Teacher Model

Model Details

Model Description

This repository contains the Marathi Teacher model developed for the research paper "Lightweight and Explainable Neural Models for Multilingual Movie Script Certification."

The model is fine-tuned from the multilingual DistilBERT architecture to classify full Marathi movie scripts into age-rating categories. It serves as the teacher model in a knowledge distillation framework used to develop lightweight student models for efficient deployment.

Developed by: Pratik N. Kalamkar
Funded by: Self-funded academic research
Shared by: Pratik N. Kalamkar
Model type: Transformer-based sequence classification model (DistilBERT)
Language(s): Marathi
License: Apache-2.0
Finetuned from model: distilbert-base-multilingual-cased

Model Sources

Uses

Direct Use

This model is intended for automated classification of Marathi movie scripts into age-rating categories.

Potential applications include:

Automated movie certification research
Movie script analysis
Content moderation
Academic NLP research
Benchmarking multilingual transformer models

Downstream Use

The model may be used as:

Teacher model for knowledge distillation
Base model for further fine-tuning
Feature extractor
Long-document text classification research

Out-of-Scope Use

This model is not intended for:

Legal or official movie certification decisions
Safety-critical systems
Hate speech detection
Sentiment analysis
Medical or legal decision making

Predictions should not replace human review.

Bias, Risks and Limitations

The model was trained using a curated dataset of Marathi movie scripts and inherits limitations of the dataset, including:

Limited dataset size
Possible class imbalance
Genre imbalance
Cultural bias
Language-specific ambiguity

Performance may decrease on scripts that differ substantially from the training distribution.

Recommendations

The model should be used as an assistive tool rather than a replacement for human reviewers.

Training Details

Training Data

The model was trained using the Marathi Movie Scripts with Verified Age Ratings https://huggingface.co/datasets/pratikkalamkar/Movie_Scripts_with_Age_Ratings_by_Pratik_Kalamkar-Mr dataset curated by the author.

Training data consists of Marathi movie scripts labelled with their corresponding age ratings.

Training Procedure

The model was fine-tuned using the Hugging Face Transformers library.

Preprocessing

UTF-8 normalization
Tokenization using DistilBERT tokenizer
Sliding-window chunking for long movie scripts
Chunk-level prediction aggregation

Training Hyperparameters

Base model: distilbert-base-multilingual-cased
Epochs: 6
Batch size (train): 8
Batch size (evaluation): 16
Learning rate: 2e-5
Weight decay: 0.01
Warmup steps: 50
Optimizer: AdamW
Maximum gradient norm: 1.0
Early stopping patience: 2
Mixed precision: Disabled (FP32)

Evaluation

Testing Data

Held-out Marathi movie script test split.

Metrics

The following metrics were used:

Accuracy
Precision
Recall
Macro F1-score
Confusion Matrix

Results

The complete experimental results are reported in the associated journal paper.

Technical Specifications

Model Architecture

DistilBERT multilingual encoder
Sequence Classification Head
Fine-tuned on Marathi movie scripts

Software

Python
PyTorch
Hugging Face Transformers
Scikit-learn

Citation

If you use this model in your research, please cite:

BibTeX

@article{kalamkar2026moviecertification,
  author = {Pratik N. Kalamkar and Prasadu Peddi and Yogesh K. Sharma},
  title = {Lightweight and Explainable Neural Models for Multilingual Movie Script Certification},
  journal = {International Journal of Information Technology and Computer Science},
  volume = {18},
  number = {2},
  pages = {146--160},
  year = {2026},
  doi = {10.5815/ijitcs.2026.02.09}
}

More Information

This model is one component of the MovieCert multilingual movie certification framework comprising:

English Teacher Model
Hindi Teacher Model
Marathi Teacher Model

The complete implementation, datasets and reproducibility materials are available through the accompanying GitHub repository and linked datasets.

Model Card Author

Pratik N. Kalamkar

Contact

For questions regarding the model or the associated research, please contact:

Pratik N. Kalamkar

GitHub: https://github.com/pratik1986

Downloads last month: 20

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for pratikkalamkar/moviecert-teacher-mr

Base model

distilbert/distilbert-base-multilingual-cased

Finetuned

(449)

this model

Dataset used to train pratikkalamkar/moviecert-teacher-mr

Collection including pratikkalamkar/moviecert-teacher-mr

Lightweight Explainable Multilingual Movie Script Certi

Collection

Official research resources accompanying the paper "Lightweight and Explainable Neural Models for Multilingual Movie Script Certification". • 6 items • Updated 15 days ago