LLama For Kathakali

A fine-tuned Llama-3.1-8B-Instruct model specialized for Kathakali ontology creation, narrative modelling, metadata generation, and domain-specific cultural reasoning.

This model assists in generating structured descriptions, story annotations, performer metadata, act interpretations, and domain-aligned outputs for classical Indian dance (Kathakali) research and digital humanities applications.

Model Details

Model Description

llama-for-kathakali is a causal language model fine-tuned on curated Kathakali-specific datasets, including:

character ontology descriptions
gestures and mudras metadata
performance sequences
story summaries (Aattakatha)
scene-level video annotations
performer attributes
domain-specific semantic structures

It supports applications such as ontology building, dataset generation, metadata expansion, and structured textual outputs for Kathakali-centered ML pipelines.

Developed by: Ashiq Firoz
Model type: Causal LM (Instruction-tuned)
Language: English (with limited Malayalam terminology support)
License: Follows upstream Meta Llama 3.1 Community License
Finetuned from: meta-llama/Llama-3.1-8B-Instruct
Access Requirement: You must have access to the base model (meta-llama/Llama-3.1-8B-Instruct) to fully use this model.

Model Sources [optional]

Repository: https://huggingface.co/Edith08/llama-for-kathakali
Paper: Not published
Demo: Not available

Uses

Direct Use

The model is designed for:

Automatic generation of Kathakali ontology entities
Metadata generation for cultural datasets
Semantic structuring of dance movements, mudras, and characters
Generating interpretive or descriptive text for research
Knowledge graph population
LLM-assisted dataset creation for downstream ML pipelines

Downstream Use

Video-to-text pipelines where generated text becomes training data
Dataset augmentation for gesture recognition or performance analysis
Text classification, extraction, and structured knowledge modeling
Fine-tuned modules for cultural heritage preservation systems

Out-of-Scope Use

The model is not intended for:

High-fidelity translation or Malayalam language generation
Medical, legal, or safety-critical decision-making
Improper cultural reinterpretation or misinformation
Generating factual claims about historical events without verification

Bias, Risks, and Limitations

Training data includes curated domain-specific texts; hence the model may exhibit:
- cultural bias towards traditional interpretations
- limited general reasoning outside Kathakali
- incomplete understanding of regional linguistic nuances
- hallucination in historical or artistic contexts
Model may over-generalize gestures, scenes, or characters if prompts are vague.

Recommendations

Use precise prompts to reduce hallucination
Prefer structured, schema-based prompts for ontology generation
Avoid using outputs as factual without expert review
Do not deploy in cultural or social decision-making contexts without human oversight

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

HF_MODEL_NAME = "Edith08/llama-for-kathakali"

tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_NAME)

model = AutoModelForCausalLM.from_pretrained(
    HF_MODEL_NAME,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "How to represent the word 'sun' in kathakali"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Note: You must have access to meta-llama/Llama-3.1-8B-Instruct to load dependency weights.

Training Details

Training Data

Training data included:

Manually curated Kathakali ontology texts
Public-domain Aattakatha literature
Annotated descriptions of mudras, characters, costumes
Semi-structured metadata designed for ontology frameworks
Additional synthetic domain-aligned text

The dataset focuses on structured and semi-structured cultural knowledge.

Training Procedure

Base Model: Llama-3.1-8B-Instruct
Fine-Tuning Method: Supervised fine-tuning (SFT)
Precision: fp16 mixed precision
Sequence length: 4096
Optimizer: AdamW
Learning rate: 2e-5
Epochs: 3–5 depending on dataset splits

Speeds, Sizes, Times [optional]

Parameter count: ~8B (inherits upstream)
Checkpoint size: ~16 GB (fp16)
Training hardware: AMD MI300X
Training duration: Approximately 6–8 hours for SFT

Evaluation

Testing Data, Factors & Metrics

Testing Data

Evaluation was conducted on withheld domain-specific texts:

Ontology concept completion
Story-element summarization
Gesture-movement consistency checks
Metadata structuring tasks

Factors

Evaluations considered:

Cultural accuracy
Ontology structure correctness
Sequence consistency
Reduction of hallucinations
Token-level coherence

Metrics

Manual qualitative evaluation
BLEU / ROUGE-L for descriptive tasks
Schema adherence score (custom heuristic)

Results

Strong performance on structured ontology generation
High consistency with Kathakali terminology
Occasional hallucinations in abstract narrative tasks
Good alignment with research workflows
Limited general conversational ability

Environmental Impact

Hardware Type: GPUs (AMD MI300X)
Hours used: ~8 hours of fine-tuning
Cloud Provider: AMD Developer Cloud
Compute Region: US-Central
Estimated CO₂ Emitted: ~15–20 kg CO₂eq (approximate)

Technical Specifications [optional]

Model Architecture and Objective

Transformer-based causal language model
32-layer architecture (inherits from Llama-3.1-8B)
Objective: next-token prediction under instruction tuning

Compute Infrastructure

Hardware: MI300X GPUs
Software:
- PyTorch
- Hugging Face Transformers
- BitsAndBytes (optional)
- Accelerate
- Python 3.10

More Information

For more information about the project or dataset creation, you may contact the author.

[More Information Needed]

Model Card Authors

Ashiq Firoz

Model Card Contact

For questions, issues, or collaboration inquiries: Email: (ashiqfiroz08@gmail.com) HF Profile: https://huggingface.co/Edith08

Downloads last month: 1