Method: TSDAE full fine-tuning (no quantization, no LoRA)
Fine-tuning: Full parameter fine-tuning (all weights updated)
Corpus: 8,464 Hindi film songs, 28,713 verse-level chunks
Scripts: Devanagari, Romanized Hindi (Hinglish), English — single unified model
Script balancing: weighted sampling to 44/44/12 (Devanagari/Hinglish/English)
Epochs: 5
Batch size: 4
Learning rate: 3e-05
Max sequence length: 400 subword tokens

Intended Use

Sentence embeddings for Hindi film song lyrics across all three script forms. Designed for thematic classification and generational/temporal analysis of Hindi music.

Limitations

Full parameter fine-tuning — no LoRA/quantization constraints on how far weights can move from the base model
No supervised thematic fine-tuning applied (Phase A only)
Not evaluated on tasks outside Hindi song lyrics

Usage

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("meet5568/bhash_finetune")

lyrics = [
    "तेरे बिना जिंदगी से कोई शिकवा नहीं",
    "tere bina zindagi se koi shikwa nahi",
    "without you life has no complaint",
]
embeddings = model.encode(lyrics)
print(embeddings.shape)  # (3, 768)