SouravAnand's picture
Update README.md
9eed26c verified
metadata
license: mit
language:
  - en
pretty_name: Synthetic On-Device Assistant Commands
size_categories:
  - 10K<n<100K
tags:
  - nlp
  - text-classification
  - synthetic-data
  - on-device-ai
  - intent-recognition
  - mobilebert

Synthetic On-Device Assistant Commands (70K Records)

Overview

This dataset, comprising 70000 unique synthetic command phrases, was created to train a robust, low-latency text classifier for an offline, private AI assistant application on Android.

It addresses the lack of publicly available, high-variability command datasets tailored for edge computing and low-latency intent recognition. The resulting model, optimized with post-training quantization, operates entirely on-device, prioritizing user privacy and speed.

Supported Tasks and Intents

The dataset is structured for a multi-class Intent Classification task across seven common mobile device actions.

Intent Category Count (Records) Example Commands
making_call 10,000 "Make a call to Anand," "Ring up my doctor."
setting_alarm 10,000 "Set alarm for 7 AM," "Wake me up at 06:45."
setting_reminder 10,000 "Remind me to pay rent tomorrow," "Set a reminder for the meeting."
playing_song 10,000 "Play 'Bohemian Rhapsody'," "Shuffle my workout playlist."
navigation 10,000 "Navigate to the nearest gas station," "Start driving home."
weather_report 10,000 "What is the weather in Croatia," "Will it rain today?"
general_question 10,000 "What's the best food in Croatia," "Tell me about the history of Rome."

Construction Methodology

The dataset was generated synthetically using BERT (MobileBERT) to create high-variability training samples.

  1. Seed Phrase Generation: 100 unique root phrases were manually created for each category (e.g., "Call [person]" or "Start navigation to [place]").
  2. Variable Injection: BERT was used to generate and inject unique, domain-specific entities (e.g., thousands of unique names for the making_call intent, and thousands of unique places for the navigation intent).
  3. Combination: This technique resulted in approximately 10,000 unique commands per category, yielding a final training set of 70,000 total records.

Data Format

The dataset is provided in a single CSV file.

Usage Notes

This dataset is perfect for training lightweight transformer models (like MobileBERT) for applications in Edge AI, on-device machine learning, and privacy-first application development.