README.md

metadata

license: cc-by-4.0
language:
  - hi
task_categories:
  - automatic-speech-recognition
  - text-generation
tags:
  - hindi
  - speech
  - instruction-following
  - llama-omni
  - indic
pretty_name: Hindi LLaMA-Omni Instruct Dataset
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - data/batch_0000[0-9][0-9]-*.parquet
          - data/batch_0001[0-9][0-9]-*.parquet
          - data/batch_000200-*.parquet
          - data/batch_000201-*.parquet
          - data/batch_000202-*.parquet
          - data/batch_000203-*.parquet
          - data/batch_000204-*.parquet
          - data/batch_000205-*.parquet
          - data/batch_000206-*.parquet
          - data/batch_000207-*.parquet
          - data/batch_000208-*.parquet
          - data/batch_000209-*.parquet
          - data/batch_000210-*.parquet
      - split: validation
        path:
          - data/batch_000211-*.parquet
          - data/batch_000212-*.parquet
          - data/batch_000213-*.parquet
          - data/batch_000214-*.parquet
          - data/batch_000215-*.parquet
          - data/batch_000216-*.parquet
          - data/batch_000217-*.parquet
          - data/batch_000218-*.parquet
          - data/batch_000219-*.parquet
          - data/batch_000220-*.parquet
          - data/batch_000221-*.parquet
          - data/batch_000222-*.parquet

Hindi LLaMA-Omni Instruct Dataset

A Hindi speech instruction-following dataset designed for training speech-language models such as LLaMA-Omni. Each example pairs a spoken Hindi user question (audio) with a text assistant response.

---

Dataset Summary

Property	Value
Language	Hindi (hi)
Total examples	~110,718
Train split	~105,000 examples (batches 001–210)
Validation split	~5,500 examples (batches 211–222)
Audio format	FLAC, 16,000 Hz mono
Conversation type	Single-turn (one user question → one assistant response)

---

Data Sources

Text conversations are sourced from ai4bharat/indic-instruct-data-v0.1, using the following subsets:

Subset	Description
Anudesh	Human-annotated Hindi instruction-response pairs
LMSYS	Chat conversations translated/adapted to Hindi
HH-RLHF	Anthropic Helpful & Harmless RLHF data in Hindi
Flan v2	Flan collection prompts in Hindi

Only single-turn conversations were retained (one user question, one assistant response).

---

Audio Generation

User questions were converted to speech using:

• TTS Model: facebook/mms-tts-hin (MMS Hindi TTS — VITS architecture)
• Sampling rate: 16,000 Hz
• Format: FLAC (lossless)

Each audio file corresponds to the user turn of a conversation and is named {id}-1_user.flac.

---

Dataset Structure

Parquet files (data/)

Each row in the parquet batches contains:

Column	Type	Description
id	string	Unique conversation ID
user_text	string	Original Hindi text of the user question
assistant_text	string	Hindi text of the assistant response
audio	Audio	FLAC audio of the user question (bytes + path)

JSON file (dataset.json)

A flat JSON array with the same conversations in message format:

{
  "id": "c01d4234-8d55-51f5-b84f-0ddfd8a271b0",
  "messages": [
    {"role": "user", "content": "न्यूयॉर्क में 3 दिवसीय यात्रा का कार्यक्रम बनाएं।"},
    {"role": "assistant", "content": "..."}
  ]
}

---

Usage

from datasets import load_dataset

ds = load_dataset("Pastaaaaa2003/hindi-llama-omni")

# Access train split
for example in ds["train"]:
    audio = example["audio"]          # dict with 'bytes' and 'path'
    question = example["user_text"]   # Hindi text
    answer = example["assistant_text"]
    print(question, "->", answer[:80])

---

License

Creative Commons Attribution 4.0 (CC BY 4.0)

Please also comply with the licenses of the original source datasets:

• ai4bharat/indic-instruct-data-v0.1
• facebook/mms-tts