10xtechnologieS/telugu_instruction_dataset

Telugu Instruction Dataset — Luuka AI

Built by 10x Technologies — a curated Telugu-language instruction-tuning dataset for training the Luuka voice assistant.

---

Overview

This dataset contains 7,496 instruction–response pairs and 399 multi-turn conversations in Telugu, covering a broad range of natural voice assistant interactions. Every response is written entirely in Telugu script — no English characters appear in any response field.

Subset	Config Name	Pairs	Description
Conversational	conversational	1,999	Greetings, small talk, everyday chitchat
Emotional Responses	emotional_responses	799	Empathetic, supportive replies
RAG	rag	1,494	Context + question → grounded answers
Safety Categories	safety_categories	599	Refusals and boundary-setting across 600 harm categories
Identity	identity	199	Luuka self-identification, company info
Multi-turn	multiturn	399	Two-turn conversations teaching context continuity
Helpful	helpful	999	Task assistance, how-to, device control
Neutral	neutral	410	Balanced, factual responses
Language Quality	language_quality	399	High-quality Telugu fluency examples
Time Context	time_context	199	Date, time, and temporal awareness

Total: ~6,496 flat pairs + 399 multi-turn conversations

---

Dataset Structure

Flat pairs (all subsets except multiturn)

{
  "instruction": "భారత రాజధాని ఏది?",
  "response": "న్యూఢిల్లీ మన భారతదేశ రాజధాని అండీ."
}

Multi-turn (multiturn subset)

{
  "turns": [
    {"instruction": "హాయ్", "response": "నమస్తే అండీ! నేను లూకా."},
    {"instruction": "మీరు ఎవరు?", "response": "నేను టెన్ ఎక్స్ టెక్నాలజీస్ తయారు చేసిన లూకాను."}
  ]
}

---

Usage

from datasets import load_dataset

# Load a specific subset
ds = load_dataset("10xtechnologieS/telugu_instruction_dataset", "conversational")
print(ds["train"][0])

# Load multi-turn conversations
ds_mt = load_dataset("10xtechnologieS/telugu_instruction_dataset", "multiturn")
print(ds_mt["train"][0])

Fine-tuning example

from datasets import load_dataset, concatenate_datasets

subsets = [
    "conversational", "emotional_responses", "rag",
    "helpful", "neutral", "language_quality",
    "identity", "time_context"
]

splits = [
    load_dataset("10xtechnologieS/telugu_instruction_dataset", s)["train"]
    for s in subsets
]

combined = concatenate_datasets(splits)
print(f"Total examples: {len(combined)}")

---

Design Principles

• Telugu-only responses — English letters (a–z, A–Z) never appear in any response field
• Voice-first phrasing — responses are natural for text-to-speech output
• Safety coverage — 600 harm categories explicitly handled with appropriate refusals
• Context continuity — multi-turn pairs teach the model to carry context across turns
• Luuka persona — consistent identity as a 10x Technologies voice assistant throughout

---

About Luuka

Luuka is a Telugu-language voice assistant developed by 10x Technologies. This dataset was created to fine-tune Luuka for natural, culturally appropriate, and safe Telugu conversations.

---

License

Apache 2.0

---

Citation

@dataset{10xtechnologies2025telugu,
  title        = {Telugu Instruction Dataset — Luuka AI},
  author       = {10x Technologies},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/10xtechnologieS/telugu_instruction_dataset}
}