Datasetpublic

adiprog14/lingrow-support-tickets

Lingrow Support Tickets (Synthetic) A synthetic dataset of 10,000 customer-support tickets for Lingrow, a real-time multilingual translation and communication platform. Each ticket contains a customer message (an error report or a how-to question), rich metadata, and a resolution. The data is fully synthetic — no real customer information is included. This dataset was built as the final project for a Data Science course. It powers the Lingrow Support Copilot: a tool that, given… See the full description on the dataset page: https://huggingface.co/datasets/adiprog14/lingrow-support-tickets.

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes239downloads

Dataset Card

Lingrow Support Tickets (Synthetic)

A synthetic dataset of 10,000 customer-support tickets for Lingrow, a real-time multilingual translation and communication platform. Each ticket contains a customer message (an error report or a how-to question), rich metadata, and a resolution. The data is fully synthetic — no real customer information is included.

This dataset was built as the final project for a Data Science course. It powers the Lingrow Support Copilot: a tool that, given a new customer message, retrieves similar past tickets and drafts a reply.

How it was created

1.Schema was derived from the official Lingrow user guide (error flows, roles, session types, audio, translation, user management).
2.A synthetic generator combined these building blocks with a hand-written phrasing bank to produce 10,000 varied tickets.
3.Text variety was improved with a Hugging Face model (humarin/chatgpt_paraphraser_on_T5_base): the original 116 base messages were paraphrased into 666 unique customer messages, lowering the duplicate-text rate from ~96.5% to ~80%.

Dataset structure

Column	Description
`ticket_id`	Unique ticket identifier (e.g. LG-100000)
`created_at`	Timestamp of the ticket
`user_role`	admin / member / guest / audience
`session_type`	quickconversation / groupconversation / live_captions
`speaking_style`	live / pushtotalk
`context_domain`	general, medical, academic, technical, legal, sport, logistics, industrial, custom
`device`	ios / android / web
`language`	Customer's language (10 options)
`error_code`	The specific error (empty for how-to questions)
`topic_key`	The topic key (error code or HOWTO_*)
`error_category`	One of 6 error categories, or `how_to`
`customer_message`	The customer's message
`resolution_steps`	The support answer
`resolved`	Whether the ticket was resolved

Exploratory Data Analysis (EDA)

—Total tickets: 10,000
—Error reports: 7,979 | How-to questions: 2,021
—Categories: 7 error categories + how-to
—Languages: 10 (balanced)
—Devices: iOS / Android / Web (balanced)
—Overall resolution rate: 88.2%
—Unique customer messages: 666

Categories and error codes are evenly distributed (a fair, balanced synthetic dataset). How-to questions are always marked resolved; error tickets are resolved ~85% of the time.

Intended use

—Information retrieval / semantic search over support tickets
—Retrieval-augmented reply drafting (RAG)
—Educational / demonstration purposes

Limitations & ethics

—Synthetic data: messages are machine-generated and may occasionally read unnaturally. They do not represent real users.
—No personal data: contains no real customer details, transcripts, credentials, or tokens.
—Error codes and flows are inspired by the Lingrow product but simplified.

License

MIT