Text Generation
Safetensors
English
Telugu
carnatic-music
konakol
solkattu
swara
tala
rhythmic-patterns
indian-classical-music
music-generation
fine-tuned
unsloth
llama
lora
Instructions to use sgattup/KonakolSwaraLLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Local Apps Settings
- Unsloth Studio
How to use sgattup/KonakolSwaraLLM with Unsloth Studio:
Install Unsloth Studio (macOS, Linux, WSL)
curl -fsSL https://unsloth.ai/install.sh | sh # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for sgattup/KonakolSwaraLLM to start chatting
Install Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for sgattup/KonakolSwaraLLM to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for sgattup/KonakolSwaraLLM to start chatting
Load model with FastModel
pip install unsloth from unsloth import FastModel model, tokenizer = FastModel.from_pretrained( model_name="sgattup/KonakolSwaraLLM", max_seq_length=2048, )
| language: | |
| - en | |
| - te | |
| license: apache-2.0 | |
| tags: | |
| - carnatic-music | |
| - konakol | |
| - solkattu | |
| - swara | |
| - tala | |
| - rhythmic-patterns | |
| - indian-classical-music | |
| - music-generation | |
| - fine-tuned | |
| - unsloth | |
| - llama | |
| - lora | |
| base_model: unsloth/llama-3.2-3B-bnb-4bit | |
| pipeline_tag: text-generation | |
| # KonakolSwara LLM | |
| **A domain-specialized generative language model for Carnatic rhythmic composition (Konakol/Solkattu) and melodic Swara sequence generation — trained to produce creative, grammatically authentic patterns in the Telugu script tradition of South Indian classical music.** | |
| --- | |
| ## Overview | |
| South Indian (Carnatic) classical music is built upon two interlocking theoretical pillars: **Tala** (rhythm) and **Raga** (melody). While much of music AI focuses on audio synthesis, this model addresses the *compositional notation* layer — the articulation of rhythmic and melodic ideas through the symbolic vocabulary that Carnatic musicians have used for centuries. | |
| **Konakol** (also called *Solkattu* in Telugu/Tamil) is the art of vocal percussion — the recitation of codified rhythmic syllables (such as *తకిట, తరికిట, తకధిమి*) that correspond precisely to strokes on the Mridangam (the double-headed barrel drum that anchors Carnatic rhythm). Konakol is simultaneously a theoretical notation system, a pedagogical tool, and a performance art in its own right. | |
| **Swara sequence composition** involves the creative arrangement of the seven swaras (Sa Ri Ga Ma Pa Da Ni — సా రి గ మ ప ద ని) into melodic phrases, scale exercises (*Varisai*), and improvisational patterns (*Kalpana Swara*). | |
| *KonakolSwara LLM* is trained to generate both categories of musical text with structural fidelity to Carnatic theory — producing compositions that are not merely plausible-looking character sequences, but reflect the internal logic of Tala cycles, Nadai (rhythmic subdivision), syllable families, and scalar grammar. | |
| --- | |
| ## Musical Background | |
| ### Konakol: Vocal Percussion | |
| Konakol syllables are derived from Sanskrit and Tamil phonemes and map to specific Mridangam strokes. The syllable families, organized by beat duration, include: | |
| | Duration | Telugu Syllables | Transliteration | | |
| |---|---|---| | |
| | 1 beat | తా, తోమ్, తహ, ధిన్ | tā, tom, tah, dhin | | |
| | 2 beats | తక, ధిమి, జిమి, ధిత్తై, గిన | taka, dhimi, jimi, dhittai, gina | | |
| | 3 beats | తకిట, తరికిట, తక్కిట, తనన | takita, tarikita, takkita, tanana | | |
| | 4 beats | తకధిమి, తకజిమి, తకజను, తరిగిన | takadhimi, takajimi, takajanu, tarigina | | |
| These syllables are combined to fill **Tala cycles** — rhythmic loops of fixed length. The most common Tala is **Adi Tala** (8 beats). The **Nadai** (gait) subdivides each beat: Chatusra (4 subdivisions), Tisra (3), Khanda (5), Misra (7). | |
| A **Korvai** is a rhythmic cadence — a phrase stated exactly three times that resolves to the first beat (Sam) of the Tala cycle, providing musical closure. | |
| ### Swara Sequences | |
| The seven swaras (Sa Ri Ga Ma Pa Da Ni), their komal/tivra variants, and their arrangements form the basis of Carnatic melodic pedagogy. Standard compositional exercises include: | |
| - **Sarali Varisai** — straight ascending/descending scale patterns in groups of four | |
| - **Janta Varisai** — doubled-note exercises (Sa Sa Ri Ri Ga Ga...) | |
| - **Alankara** — structured patterns in different speeds and Nadais | |
| - **Kalpana Swara** — free melodic improvisation within a raga's grammatical constraints | |
| --- | |
| ## Capabilities | |
| This model can: | |
| - **Generate complete Konakol compositions** in a specified Tala and Nadai, using authentic syllable combinations | |
| - **Compose Korvai patterns** — three-fold resolution phrases with correct Sam landing | |
| - **Generate Swara sequence compositions** in ascending, descending, zigzag, and raga-specific patterns | |
| - **Demonstrate Kuraippu** (gradual reduction technique where phrases shorten over successive repetitions) | |
| - **Demonstrate Nadai Bheda** (switching rhythmic gait within a single composition) | |
| - **Explain Konakol theory** — syllable families, beat values, Tala structures, Nadai theory | |
| - **Explain Swara theory** — Varisai types, Alankara, Kalai (speed), Kalpana Swara | |
| - **Generate combined Konakol + Swara compositions** that pair rhythmic and melodic notation | |
| --- | |
| ## How to Use | |
| ```python | |
| from transformers import AutoModelForCausalLM, AutoTokenizer | |
| import torch | |
| model_id = "sgattup/KonakolSwaraLLM" | |
| tokenizer = AutoTokenizer.from_pretrained(model_id) | |
| model = AutoModelForCausalLM.from_pretrained( | |
| model_id, | |
| torch_dtype=torch.float16, | |
| device_map="auto" | |
| ) | |
| PROMPT = """You are an expert in Carnatic classical music, specializing in Konakol (Solkattu) — the vocal recitation of rhythmic syllables — and Swara sequence composition. You can explain Tala theory, compose creative Konakol patterns, generate melodic Swara sequences, and teach the grammar of South Indian rhythm and melody. | |
| ### Question: | |
| {} | |
| ### Answer: | |
| """ | |
| def ask(question, max_tokens=800): | |
| prompt = PROMPT.format(question) | |
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) | |
| outputs = model.generate( | |
| **inputs, | |
| max_new_tokens=max_tokens, | |
| temperature=0.8, | |
| do_sample=True | |
| ) | |
| response = tokenizer.decode(outputs[0], skip_special_tokens=True) | |
| return response.split("### Answer:")[-1].strip() | |
| # Generate a Konakol composition | |
| print(ask("Compose a Konakol pattern in Adi Tala using takita and tarikita combinations.")) | |
| # Generate swara sequences | |
| print(ask("Generate a creative swara composition with all seven notes in Telugu notation.")) | |
| # Ask about theory | |
| print(ask("What is a Korvai and how is it constructed?")) | |
| # Generate a Korvai | |
| print(ask("Create a Korvai — a three-fold resolution pattern that lands on Sam in Adi Tala.")) | |
| ``` | |
| --- | |
| ## Illustrative Outputs | |
| **Q: Compose a Konakol pattern in Adi Tala, Chatusra Nadai:** | |
| > తకిట తరికిట తకిట తరికిట | |
| > తక్కిట తక్కిట తోమ్ తా | |
| > తకధిమి తకజిమి తోమ్ తోమ్ | |
| > తరిగిన తకజను తా | |
| > ధిత్తై తహ తకిట తహ | |
| > ధిమితక జిమితక తోమ్ తా | |
| **Q: Generate a Sarali Varisai swara pattern:** | |
| > సా రి గ మ / ప ద ని సా / సా ని ద ప / మ గ రి సా | |
| > రి గ మ ప / ద ని సా రి / రి సా ని ద / ప మ గ రి | |
| > గ మ ప ద / ని సా రి గ / గ రి సా ని / ద ప మ గ... | |
| **Q: What is Kuraippu?** | |
| > Kuraippu (from Tamil/Telugu కురైప్పు) is a rhythmic device in Carnatic Konakol and Mridangam solos where a phrase is stated in full, then restated in progressively shorter versions — each iteration trimming syllables from the beginning or end while maintaining the rhythmic alignment with the Tala cycle. The effect is one of musical intensification and condensation: the pattern feels like it is being compressed toward a point of culmination. A Kuraippu typically concludes on Sam (the first beat), making it a variant of the Korvai principle... | |
| --- | |
| ## Model Specifications | |
| | Parameter | Value | | |
| |---|---| | |
| | Base Model | LLaMA 3.2 3B (`unsloth/llama-3.2-3B-bnb-4bit`) | | |
| | Fine-tuning Method | QLoRA via Unsloth | | |
| | Quantization | 4-bit NF4 with double quantization | | |
| | LoRA Rank | 16 | | |
| | LoRA Alpha | 32 | | |
| | Training Epochs | 3 | | |
| | Learning Rate | 2e-4 | | |
| | Training Hardware | Google Colab T4 GPU (free tier) | | |
| | Dataset | 25+ Konakol compositions + 15+ Swara compositions + 15+ theory pairs | | |
| | Script Languages | Telugu (తెలుగు) and English | | |
| | License | Apache 2.0 | | |
| --- | |
| ## Limitations | |
| - **Tala mathematics:** The model generates syllable patterns that are structurally authentic in style; however, beat-count precision is not mathematically guaranteed. A practicing musician should verify rhythmic sums before performance use. | |
| - **Raga specificity:** Swara sequences are generated based on general scalar logic. For raga-specific composition, pair this model with [RagaLakshanaLLM](https://huggingface.co/sgattup/RagaLakshanaLLM). | |
| - **Telugu script rendering:** Proper display requires Telugu Unicode font support in the rendering environment. | |
| - **Creative vs. canonical:** Generated Konakol compositions are creative/novel, not transcriptions of traditional compositions. | |
| --- | |
| ## Companion Models | |
| - **[sgattup/RagaLakshanaLLM](https://huggingface.co/sgattup/RagaLakshanaLLM)** — Raga theory, Lakshana, Thaat/Melakarta classification | |
| - **[sgattup/IndianCultureLLM](https://huggingface.co/sgattup/IndianCultureLLM)** — Broad Indian culture, history, philosophy, and arts | |
| --- | |
| ## Training Code | |
| [github.com/sai-educ/indian-culture-llm](https://github.com/sai-educ/indian-culture-llm) | |
| --- | |
| ## Citation | |
| ```bibtex | |
| @misc{KonakolSwaraLLM2026, | |
| author = {sgattup}, | |
| title = {KonakolSwara LLM: A Generative Model for Carnatic Rhythmic and Melodic Composition}, | |
| year = {2026}, | |
| publisher = {HuggingFace}, | |
| howpublished = {\url{https://huggingface.co/sgattup/KonakolSwaraLLM}} | |
| } | |
| ``` | |
| --- | |
| *This model is part of a series of specialized AI models for Indian classical music and culture, developed to make the theoretical and creative vocabulary of Indian traditions more accessible through natural language interfaces.* | |