---
license: mit
language:
  - ps
  - fa
  - ur
  - bal
  - snd
  - en
  - ja
tags:
  - restaurant-chatbot
  - reasoning
  - multilingual
  - low-resource-languages
  - pashto
  - farsi
  - dari
  - urdu
  - balochi
  - sindhi
  - synthetic-data
  - SFT
  - conversational-ai
task_categories:
  - text-generation
  - question-answering
---

# Polyglot-Restaurant-Reasoning-Dataset 🍽️🌐🤖

## Dataset Summary
The **Polyglot-Restaurant-Reasoning-Dataset** is a specialized instruction-tuning dataset engineered to train large language models (LLMs) to function as intelligent, context-aware, and reasoning-capable restaurant chatbot assistants. 

This dataset specifically focuses on empowering **low-resource and regional languages** (**Pashto, Farsi/Dari, Urdu, Balochi, and Sindhi**), alongside bridging support for **English** and **Japanese**. 

Unlike standard conversational corpora, this dataset embeds explicit internal **reasoning traces** (chain-of-thought) across these linguistic landscapes, teaching models how to dynamically process table bookings, complex dietary constraints, and regional hospitality nuances before generating the final customer-facing response.

---

## Repository Structure & Files
The dataset is organized into separate JSON Lines (`.jsonl`) files by language under the `data/` directory:

- `data/pashto_responses_1.jsonl` (Pashto - `ps`)
- `data/farsi_responses_1.jsonl` (Farsi/Dari - `fa`)
- `data/urdu_responses_1.jsonl` (Urdu - `ur`)
- `data/balochi_responses_1.jsonl` (Balochi - `bal`)
- `data/sindhi_responses_1.jsonl` (Sindhi - `snd`)
- `data/english_responses_1.jsonl` (English - `en`)
- `data/japanese_responses_1.jsonl` (Japanese - `ja`)

---

## Dataset Schema

### Data Fields
- `id`: Unique string identifier for the interaction.
- `language`: The ISO language code (e.g., `ps`, `fa`, `ur`, `bal`, `snd`, `en`, `ja`).
- `prompt`: The incoming customer query or instruction regarding reservations, menu inquiries, or dietary needs.
- `reasoning_trace`: The hidden step-by-step logic used by the model to deduce constraints, verify policies, and plan the response.
- `response`: The final polished, hospitable response delivered to the customer in the target language.

### Data Sample Example (Pashto - `ps`)
```json
{
  "id": "rest_reason_ps_001",
  "language": "ps",
  "prompt": "سلام، ایا د جمعې په ورځ ماښام په ۷ بجو د ۶ کسانو لپاره میز شته؟ زموږ سره یو ملګری دی چې د ځمکنیو مرچو (peanuts) کلکه الرجی لري.",
  "reasoning_trace": "1. Analyze user request: Party of 6, Friday at 7 PM, severe peanut allergy.\n2. Check constraints: Group reservation requires confirmation. Allergy is critical for kitchen safety.\n3. Formulate plan: Welcome the customer, confirm table availability, assure them that the kitchen will take strict precautions regarding the peanut allergy.",
  "response": "سلام! موږ به ډیر خوښ شو چې د جمعې په ورځ ماښام په ۷ بجو ستاسو ۶ کسانو کوربتوب وکړو. ستاسو د ملګري د الرجی په اړه اندیښنه مه کوئ؛ زموږ پخلنځی کلک احتیاط کوي او موږ به پخلنځي ته لارښوونه وکړو چې د کراس کانټیمینشن مخنیوی وکړي. ایا غواړئ چې زه دا ریزرویشن تایید کړم؟"
}

```

---

## Supported Tasks & Use Cases

* **Supervised Fine-Tuning (SFT):** Fine-tuning open-source models for domain-specific, localized conversational agents.
* **Low-Resource Chain-of-Thought (CoT):** Advancing multi-step logic and reasoning capabilities in under-resourced language models.
* **Multilingual Assistant Deployment:** Building culturally nuanced and hospitable virtual assistants across South/Central Asian and international markets.

---

## Licensing Information

This dataset is released under the **MIT License**.

```

```