เฐจเฐพเฐฆเฐฟเฐจเฐฟ (Naadini) - Telugu Creative AI

เฐจเฐพเฐฆเฐฟเฐจเฐฟ means "She who is the source of divine sound, poetry and wisdom"


Latest - V4 (Balanced Specialist)

V4 builds on V3's strong poetry foundation and adds much better songs and stories!

  • Resumed training from V3 (poetry specialist)
  • Added curated high-quality songs with proper pallavi/charanam structure
  • Added curated stories with clear narrative and moral
  • Kept poetry quality while broadening other skills

Version History & Comparison

Feature V1 V2 V3 V4 (Current)
Base Model Gemma 2B Qwen3 14B Qwen3 14B Qwen3 14B
Val Loss 2.60 0.493 0.208 0.678
Training Focus Mixed Mixed Poetry only Poetry + Songs + Stories
Poetry Quality Basic Good Excellent Excellent
Songs Poor Good Okay Excellent
Stories Poor Okay Okay Much Better
Conversations Good Good Good Good
Overall Balance Poor Okay Specialist Best Balanced

Note: Val loss for V4 is measured on a broader, mixed validation set (poetry + songs + stories + general), so it is not directly comparable to V3's poetry-only val loss. V4 is the most well-rounded version.

V4 is our best all-round model!


Capabilities

Capability Quality
Telugu Poems (เฐ•เฐตเฐฟเฐคเฐฒเฑ) EXCELLENT
Classical Padyalu (เฐชเฐฆเฑเฐฏเฐพเฐฒเฑ) EXCELLENT
Telugu Songs with proper structure EXCELLENT
Devotional Songs (เฐญเฐ•เฑเฐคเฐฟ เฐชเฐพเฐŸเฐฒเฑ) EXCELLENT
Folk Songs (เฐœเฐพเฐจเฐชเฐฆ เฐชเฐพเฐŸเฐฒเฑ) EXCELLENT
Telugu Stories with narrative GOOD
Telugu Conversations GOOD
Mother Songs (เฐ…เฐฎเฑเฐฎ เฐชเฐพเฐŸเฐฒเฑ) EXCELLENT

Sample Outputs from V4

Folk Song (เฐœเฐพเฐจเฐชเฐฆ เฐชเฐพเฐŸ)

เฐชเฐฒเฑเฐฒเฐตเฐฟ: เฐชเฐฒเฑเฐฒเฑ† เฐœเฑ€เฐตเฐฟเฐคเฐ‚ เฐชเฑ†เฐฆเฑเฐฆ เฐชเฑเฐฐเฑ‡เฐฎ, เฐจเฑ€เฐŸเฐฟ เฐฎเฑ€เฐฆ เฐ•เฑเฐฒเฑเฐฒเฑ เฐชเฐพเฐŸเฑ‡ เฐ—เฐพเฐจเฐ‚. เฐชเฐฟเฐฒเฑเฐฒเฐฒ เฐจเฐตเฑเฐตเฑ เฐชเฐฒเฑเฐฒเฐตเฐฟ เฐฒเฐพเฐ‚เฐŸเฐฟเฐฆเฐฟ, เฐฎเฐพเฐฎเฐฟเฐกเฐฟ เฐตเฑƒเฐ•เฑเฐทเฐ‚ เฐšเฑ†เฐŸเฑเฐŸเฑ เฐ•เฐพเฐฒเฐ‚.

เฐšเฐฐเฐฃเฐ‚ 1: เฐชเฐฒเฑเฐฒเฑ† เฐ—เฑเฐกเฐฟ เฐšเฑ‚เฐกเฐพเฐฒเฐ‚เฐŸเฑ‡ เฐฎเฐจเฐธเฑเฐธเฑ เฐชเฐพเฐŸเฑ, เฐ•เฑ‹เฐกเฐฟ เฐจเฐฟเฐฆเฑเฐฐ เฐ—เฑเฐ‚เฐกเฑ† เฐ•เฑ‚เฐกเฐพ เฐชเฐพเฐŸเฑ.

Devotional (เฐญเฐ•เฑเฐคเฐฟ เฐชเฐพเฐŸ)

เฐชเฐฒเฑเฐฒเฐตเฐฟ: เฐตเฑ†เฐ‚เฐ•เฐŸเฑ‡เฐถเฑเฐตเฐฐเฐพ เฐจเฑ€ เฐชเฑ‡เฐฐเฑ เฐตเฐฟเฐจเฑเฐจ เฐจเฐพเฐ•เฑ เฐŠเฐชเฐฟเฐฐเฐฟ เฐคเฑ€เฐธเฑเฐ•เฑ‹เฐฒเฑ‡เฐ•เฐชเฑ‹เฐคเฑเฐ‚เฐฆเฐฟ เฐจเฑ€ เฐญเฐ•เฑเฐคเฑเฐฒ เฐฎเฐจเฐธเฑเฐฒเฑ‹ เฐจเฐฟเฐฒเฐฟเฐšเฐฟเฐจ เฐจเฑ€ เฐชเฑ‡เฐฐเฑ เฐ…เฐฎเฑƒเฐคเฐฎเฑเฐตเฐฒเฑ‡ เฐจเฐพ เฐฎเฐจเฐธเฑเฐจเฑ เฐจเฐฟเฐ‚เฐชเฑเฐคเฑเฐ‚เฐฆเฐฟ

Mother Song (เฐ…เฐฎเฑเฐฎ เฐชเฐพเฐŸ)

เฐ…เฐฎเฑเฐฎเฐพ เฐ…เฐฎเฑเฐฎเฐพ เฐจเฑ€ เฐชเฑ‡เฐฐเฑ เฐจเฐพ เฐœเฑ€เฐตเฐฟเฐคเฐ‚เฐฒเฑ‹ เฐฎเฑ†เฐฒเฑเฐฒเฐ—เฐพ เฐจเฐพ เฐ—เฑเฐ‚เฐกเฑ†เฐฒเฑ‹ เฐจเฐฟเฐฒเฐฌเฐกเฐฟเฐ‚เฐฆเฐฟ เฐจเฑ€ เฐšเฑ‚เฐชเฑเฐฒเฑเฐฒเฑ‹ เฐชเฑเฐฐเฑ‡เฐฎ เฐจเฐพ เฐœเฑ€เฐตเฐฟเฐคเฐพเฐจเฑเฐจเฐฟ เฐฎเฐพเฐฐเฑเฐšเฐฟเฐ‚เฐฆเฐฟ

เฐ•เฑ‹เฐฐเฐธเฑ: เฐ…เฐฎเฑเฐฎเฐพ เฐจเฐฟเฐจเฑเฐจเฑ เฐจเฑ‡เฐจเฑ เฐชเฑเฐฐเฑ‡เฐฎเฐฟเฐธเฑเฐคเฐพเฐจเฑ เฐจเฑ€ เฐชเฑเฐฐเฑ‡เฐฎ เฐจเฐพ เฐœเฑ€เฐตเฐฟเฐคเฐ‚


Training Data

Real Classical Telugu Literature (Wikisource)

  • เฐตเฑ‡เฐฎเฐจ เฐชเฐฆเฑเฐฏเฐพเฐฒเฑ, เฐธเฑเฐฎเฐคเฐฟ เฐถเฐคเฐ•เฐ‚, เฐ•เฐพเฐณเฐนเฐธเฑเฐคเฑ€เฐถเฑเฐตเฐฐ เฐถเฐคเฐ•เฐ‚
  • เฐ…เฐจเฑเฐจเฐฎเฐพเฐšเฐพเฐฐเฑเฐฏ เฐ•เฑ€เฐฐเฑเฐคเฐจเฐฒเฑ, เฐคเฑเฐฏเฐพเฐ—เฐฐเฐพเฐœ เฐ•เฑ€เฐฐเฑเฐคเฐจเฐฒเฑ, เฐฐเฐพเฐฎเฐฆเฐพเฐธเฑ เฐ•เฑ€เฐฐเฑเฐคเฐจเฐฒเฑ
  • เฐชเฑ‹เฐคเฐจ เฐญเฐพเฐ—เฐตเฐคเฐ‚, เฐฎเฐจเฑเฐšเฐฐเฐฟเฐคเฑเฐฐ, เฐ†เฐฎเฑเฐ•เฑเฐคเฐฎเฐพเฐฒเฑเฐฏเฐฆ

Curated Generated Data (V4 addition)

  • 50 high-quality Telugu songs with proper structure
  • 50 high-quality Telugu stories with narrative and moral
  • 100 classical padyalu and 100 modern kavithalu

General Telugu Data

  • Telugu instruction datasets sample
  • Telugu conversations and instructions

Training Details

Detail Value
Base Model Qwen3-14B (4-bit quantized)
Training Framework MLX LoRA
Hardware Apple M4 Mac Mini (24GB)
V4 Iterations 3,000 (resumed from V3)
V4 Train Loss 0.593
V4 Val Loss 0.678 (broader mixed validation set)
V4 Training Data 4,323 curated examples
Num Layers 16
Learning Rate 5e-6
Batch Size 2

How to Use

pip install mlx-lm

from mlx_lm import load, generate from mlx_lm.sample_utils import make_sampler

model, tokenizer = load( "mlx-community/Qwen3-14B-4bit", adapter_path="mohantvj/naadini-telugu" )

messages = [ { "role": "system", "content": "You are naadini, a Telugu creative writing assistant. Write in Telugu script. /no_think" }, { "role": "user", "content": "varsham gurinchi oka andhamaina kavita raayandi /no_think" } ]

text = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=False ) text = text + "

"

sampler = make_sampler(temp=0.7, top_p=0.9, min_p=0.1) response = generate( model, tokenizer, prompt=text, max_tokens=400, sampler=sampler, verbose=True ) print(response)


Requirements

  • Apple Silicon Mac (M1/M2/M3/M4)
  • 16GB+ RAM recommended (24GB ideal)
  • Python 3.9+
  • mlx-lm library

Creator

Made with love by Thonangi Venkata Jagan Mohan

For Telugu community - telugu vaarandhariki ankitam

HuggingFace: https://huggingface.co/mohantvj


License

Apache 2.0 - Free to use, modify and build upon!

Acknowledgements

  • MLX Community for Qwen3-14B-4bit
  • Telugu Wikisource for classical literature
  • AI4Bharat for Telugu datasets
  • Apple for MLX framework
  • Telugu community for downloads and support!
Downloads last month
18
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ 1 Ask for provider support

Model tree for mohantvj/naadini-telugu

Finetuned
Qwen/Qwen3-14B
Adapter
(2)
this model

Space using mohantvj/naadini-telugu 1