We've released Poetry Llama, a 70 billion parameter language model fine-tuned on classical Urdu poetry. This represents months of work in bringing together modern AI with one of the richest poetic traditions.

Why Urdu Poetry?

Urdu poetry, with its rich traditions of ghazals, nazms, and shers, represents centuries of literary excellence. Poets like Ghalib, Mir, Faiz, and Iqbal have created works that continue to resonate deeply with millions.

However, existing language models struggle with:

  • The nuanced metaphors and imagery
  • Classical poetic structures and meters
  • Proper understanding of cultural context
  • The elegant Nastaliq script

The Model

Poetry Llama is built on Llama 3.3 70B and fine-tuned on:

  • 10,000 classical shers (couplets)
  • 25,000 complete ghazals
  • Multiple eras and poetic styles
  • Verified and validated content

Capabilities

The model can:

  • Generate poetry in classical styles
  • Complete incomplete verses
  • Explain poetic metaphors and meanings
  • Maintain proper meter and rhythm
  • Work with both Roman and Nastaliq scripts

Technical Details

  • Base Model: Llama 3.3 70B
  • Training Tokens: ~7M
  • Fine-tuning Approach: LoRA with custom tokenization
  • License: Llama license
  • Hardware: Trained on L40S GPUs

Try It Out

The model is available on Hugging Face. We're also working on a web demo for easy access.

Future Work

We're exploring:

  • Expanding to other Urdu literary forms
  • Multi-lingual poetry models
  • Educational applications