HN Reader

I’m excited to share NepaliGPT, an open-source language model tailored for both Nepali and English, finetuned on high-quality Nepali datasets and built on Llama 3.1-8B-Instruct. It’s the first robust, freely available generative AI designed for Nepali, released under the MIT license.

Why?

Nepali is a low-resource language in the NLP world: foundational tools and models are scarce, and existing solutions are often closed or commercial. I built NepaliGPT to fill this gap and support open research, local application development, and broader innovation.

How does it work?

Finetuned with Hugging Face TRL & Unsloth for efficiency

Supports Nepali and English for tasks like translation, summarization, code generation, and Q&A

Built on Llama 3.1-8B-Instruct for strong baseline compatibility and performance

Features:

1.MIT license: completely free for research and commercial use

2.Easy integration: Hugging Face compatible, safetensors available.

3.Excel in Nepali language than exiting models.

4.Sample outputs and usage docs included.

Collaborate:

I’m looking for feedback from linguists, devs, educators, and AI researchers

Know of high-quality Nepali datasets or have real-world use cases? Let’s connect

Contributors welcome for evaluations, dataset improvements, or downstream apps

I hope NepaliGPT helps empower the open-source NLP ecosystem for Nepali and inspires similar work for low-resource languages. Looking forward to your feedback, questions, and ideas for partnership!