Large language models have come a long way. But if you actually speak Gujarati, there's something you notice pretty quickly.
Ask many open-source models a question in Gujarati, and you get one of a few things back: a mix of languages, phrasing that no native speaker would ever use, or sometimes the model just gives up and replies in Hindi or English. It's frustrating, especially when you know the technology clearly exists — it just hasn't been pointed in this direction.
I started looking for quality instruction-tuning datasets for Gujarati, thinking someone must have built something solid by now. There wasn't much. Most open datasets focus on high-resource languages, and Gujarati — with over 60 million speakers — was mostly left out.
So I built one.
What I Built
The goal wasn't to create the biggest dataset out there. It was to create something actually useful: a general-purpose Gujarati instruction dataset for fine-tuning, benchmarking, and experimenting with Gujarati-capable language models.
The result is GGJI v1 (Gujarati General-Purpose Instruction Dataset) — 23,181 instruction-response pairs across 18 behavioral task categories:
- General and explanatory question answering
- Multi-step reasoning
- Summarization
- Translation (English ↔ Gujarati)
- Grammar and spelling correction
- Paraphrasing and rewriting
- Creative writing (stories, poems, dialogues)
- Conversational / dialogue assistant
- Code explanation in Gujarati
- Culturally grounded content (Gujarat history, customs, traditions)
- Structured and formatted output
- Safety and refusal handling
- And more
The intent behind covering so many categories is to give models exposure to diverse instruction formats, rather than the same type of question-answer pattern repeated thousands of times. A fine-tuned model should be able to follow instructions in Gujarati, not just look things up.
Why This Matters
Instruction tuning has become one of the key techniques for making language models useful in real conversations. Good instruction datasets help models understand user intent, generalize across task types, and produce responses that feel natural rather than robotic.
For English, there are hundreds of public datasets to pick from. For Gujarati, the ecosystem is still in its early stages.
I'm not claiming this dataset solves everything. But 60 million people deserve better tools in their own language, and the gap won't close unless people actually start building. This is my attempt to contribute something concrete.
Open Source, No Strings Attached
The entire dataset — all 23,181 records — is publicly available on Hugging Face under the Apache 2.0 license. You can use it for research, model development, commercial projects, or personal experiments.
🔗 Dataset: https://huggingface.co/datasets/tkdonda/gujarati-general-purpose-instruction
What's Next
This is v1. There are gaps I already know about — some dialects (Kathiawadi, Surti) aren't well represented, and a few categories are thinner than I'd like. I plan to keep improving it based on feedback and usage.
If you're working on Gujarati NLP, multilingual LLMs, or low-resource language AI in general, I'd love to hear from you. Suggestions, issues, and contributions are always welcome.
Building better AI for underserved languages is something the community can only do together. Hopefully this is a useful starting point.
Top comments (0)