meet_models: Hindi song theme classification

How to access and run the models: see HOW_TO_USE.md. Quick start:

huggingface-cli login
huggingface-cli download meet5568/meet_models --local-dir meet_models
cd meet_models/code && pip install -r requirements.txt
python predict.py --model-dir .. --lyrics "tujhe dekha toh yeh jaana sanam"

All models of the project Multimodal Thematic Classification and Generational Analysis of Hindi Songs: classifying Hindi film songs as romance, sadness, celebration or motivational from their lyrics, singing voice and music.

lyrics_encoder/
  tsdae/                       BhashaEmbed adapted to song lyrics with TSDAE (no labels)
  theme_tuned/fold_0 ... 4/    + triplet fine-tuning on the four themes, one encoder per CV fold
  theme_tuned/training_summary.json, training_log.json
classification/
  <combination>/fold_<k>/      chosen classifier per input combination and fold (75)
  index.csv, README.md
code/                          predict.py (run the models), classify.py (model definitions),
                               config.yaml, save_models.py, requirements.txt
HOW_TO_USE.md                  access, download, run

Results (5-fold cross-validated test macro-F1; 0.25 = chance)

input	macro-F1
audio only (vocal + instrument + mix)	0.513
lyrics, TSDAE encoder	0.612
lyrics, theme-tuned encoder	0.682
lyrics + instrument + mix (best)	0.699
average of the 8 lyrics-based models	0.705

Important: encoder theme_tuned/fold_k and classifiers fold_k were trained without the songs of test fold k. Use the fold-k models only to evaluate on fold k; for new songs, average the predictions of the five folds.

Base model: AkshitaS/bhasha-embed-v0 (MuRIL-based, Devanagari Hindi, romanized Hindi and English). Full description: the project report.