---
license: other
language: [hi]
tags: [hindi, instruction-tuning, indic, recipe, ai4bharat]
---

# 🇮🇳 Hindi Instruct Mixed (6k) — Build Recipe

The **script** to reproduce a 6k cleaned Hindi instruction set for fine-tuning small Hindi LLMs.
It does **not** re-host data — run `build.py` to generate it locally from AI4Bharat source.

## Recipe
- **Source:** AI4Bharat `indic-instruct-data-v0.1` — `anudesh` + `dolly` (Hindi)
- **Filter:** dolly kept at **chrF ≥ 50** (drops weak machine-translations; 14,928 / 15,011)
- **Balance:** ~3k per source → 6,000 · output `hindi_instruct_mixed_6k.jsonl` (role/content)

## ⚠️ Licensing / provenance
The underlying data is **not uniformly licensed** — only the script is shared here:
- **Dolly** portion → Databricks Dolly lineage (**CC-BY-SA-3.0**)
- **Anudesh** responses → generated by **Llama 2 70B** (**Llama 2 license**, use restrictions)
- Translation: IndicTrans2 (AI4Bharat)

Review AI4Bharat's [dataset card](https://huggingface.co/datasets/ai4bharat/indic-instruct-data-v0.1)
and the [Airavata report](https://arxiv.org/abs/2401.15006) for full terms.
Credit: **AI4Bharat**. This repo adds only filtering, balancing, and reformatting logic.