sainitishb/Kavya-1-7B
sainitishb/Kavya-1-7B is a 7.6 billion parameter Telugu lyric-writing model, fine-tuned from Qwen2.5-7B-Instruct, that composes original songs in traditional pallavi–charanam form. Trained on 9,100 curated Telugu songs, it excels at generating lyrics with the imagery, emotional register, and metrical rhythm of modern Telugu film music. This model is specifically designed for creative Telugu songwriting, offering a unique capability in a domain where general-purpose models often produce stilted output.
Loading preview...
Overview
Kavya-1-7B is a 7.6 billion parameter model developed by sainitishb, specifically designed for composing original Telugu song lyrics. Fine-tuned from Qwen2.5-7B-Instruct, it generates complete songs structured as pallavi, anupallavi, and charanam, capturing the imagery, emotional register, and metrical rhythm of modern Telugu film music.
Key Capabilities
- Telugu Song Composition: Generates original lyrics in Telugu script based on provided themes, moods, and styles.
- Poetic Form Adherence: Composes songs adhering to traditional Telugu poetic structures like pallavi–charanam.
- Culturally Grounded Output: Trained on 9,100 curated Telugu songs, blending contemporary cinema and classical devotional styles.
Intended Use Cases
- Creative Songwriting: Ideal for drafting and ideating Telugu song lyrics.
- Songwriting Assistance: Helps in generating alternate refrains, verses, or imagery.
- Research: Useful for studies in low-resource language generation and Indic poetic forms.
Known Limitations
- Orthography Issues: Due to the base model's tokenizer lacking Telugu characters, generated output frequently contains malformed words and broken conjuncts. This is a significant defect, making it a research release rather than a production-ready model.
- Telugu Script Only: Performance degrades significantly with Romanized Telugu input.
- Film-Song Register: Primarily excels in contemporary cinema style; other registers are weaker.
- Long Generation Drift: Thematic coherence can degrade beyond approximately 650 tokens.