KavithaSaaram (9b) Model Card

KavithaSaaram-9b-it is a fine-tuned large language model built for intralingual translation of classical Tamil poetry into modern, understandable Tamil. It bridges the linguistic gap between Sangam and classical Tamil literature and today's Tamil speakers, with a focus on poetic and cultural preservation.

Model Details

Model Description

KavithaSaaram-9b-it is a 9-billion parameter decoder-only language model fine-tuned for intralingual poetry translation. The model is part of the PaaPeyarchi project, a research initiative to simplify ancient Tamil literary texts using modern Tamil for educational, literary, and cultural applications. It leverages a curated dataset of classical Tamil poems paired with simplified, culturally faithful translations.

Developed by: Diwahar A K (@akdiwahar)
Shared by: Diwahar A K
Model type: Decoder-only language model (LLM). Foundation Model is Gemma 2 9b instruct Model
Language(s): Tamil
Finetuned from model: google/gemma-2-9b-it
License: gemma
Finetuned from dataset: akdiwahar/PaaPeyarchi

Model Sources

Repository: https://huggingface.co/akdiwahar/KavithaSaaram-9b-it
Paper: Submission in progress to IconDeepCom 2025, track: Real-World Applications of AI

Results

KavithaSaaram-9b-it shows significant improvements over KavithaSaaram-2b across all major metrics.
The 9B model produces translations that are more fluent, faithful, and easier to understand while preserving poetic structure.

Metric	KavithaSaaram 2B	KavithaSaaram 9B	What is Good	Improvement
BLEU	0.118	0.382	Higher	224% better
METEOR	0.299	0.619	Higher	107% better
CHRF	0.424	0.639	Higher	51% better
GLEU	0.184	0.486	Higher	164% better
TER	0.953	0.585	Lower	39% better
WER	0.877	0.480	Lower	45% better

Citation

If you use this model in your research, please cite:

APA:

Diwahar A.K.(2025). KavithaSaaram-9b-it: A Large Language Model for Intralingual Tamil Poetry Translation [Computer software]. Hugging Face. https://huggingface.co/akdiwahar/KavithaSaaram-9b-it

BibTeX:

@misc{akdiwahar2025kavithasaaram9b,
  title        = {KavithaSaaram-9b-it: A Large Language Model for Intralingual Tamil Poetry Translation},
  author       = {Diwahar A.K.},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/akdiwahar/KavithaSaaram-9b-it}},
  note         = {Fine-tuned LLM for simplifying classical Tamil poetry}
}

Uses

Direct Use

This model is intended for the intralingual translation of Tamil poetry, particularly:

Translating Tamil poems into simplified modern Tamil.
Preserving poetic structures while enhancing comprehension.
Supporting students, educators, and researchers in Tamil literary studies.

Downstream Use

Can be integrated into educational tools and apps.
Suitable for research in linguistic preservation and low-resource language translation.
Useful in digital humanities and AI-based cultural preservation projects.

Out-of-Scope Use

Translation to or from non-Tamil languages.
Use in legal, medical, or high-stakes decision-making contexts.
Direct generation of modern creative content outside the poetic translation domain.

Bias, Risks, and Limitations

The model might:

Over-simplify culturally dense or metaphorical phrases.
Misinterpret archaic grammatical constructs.
Occasionally introduce hallucinations or modern idioms not present in the original poem.

Recommendations

Use alongside human expert verification for academic or publishing purposes.
Consider cross-checking with multiple classical commentaries where applicable.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("akdiwahar/KavithaSaaram-9b-it")
model = AutoModelForCausalLM.from_pretrained("akdiwahar/KavithaSaaram-9b-it")

prompt = "Tamil Intralingual Translation: 'அன்பிற்கும் உண்டோ அடைக்குந்தாழ் ஆர்வலர் புன்கணீர் பூசல் தரும்'"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512,repetition_penalty=1.1,no_repeat_ngram_size=4,temperature=1, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

akdiwahar
/

KavithaSaaram-9b-it