Support ticket dataset spec

Why tickets are close to labelled tasks

Most enterprise text records what people said. A ticket also records what happened next: who picked it up, which category it was filed under, how long it waited, whether it was solved, and whether it came back. That ending is the reason tickets are worth specifying carefully. With the outcome attached, a conversation becomes a task with a result, which is what evaluation and reinforcement learning need.

The schema keeps the customer-facing exchange and the internal notes in one ordered list, separated by the visibility field. Notes are often the most instructive part: they show the reasoning an agent did not write to the customer, the escalation to a colleague and the workaround that was tried first. Status and field changes come in as events between turns, so you can reconstruct the queue history without a second file.

Outcome codes need a code book

Every helpdesk lets an organisation invent its own categories and resolution codes, and those labels drift over the years. A dataset that ships codes without explaining them is hard to use. Ask for a code book in the datasheet that maps each code to a plain description, the period it was in use, and how often it appears. If codes changed mid-period, ask for a mapping or for the period to be split.

Building tasks from tickets

For RL, the useful subset is tickets where the resolution can be checked: a refund issued, a setting changed, an article linked, a bug confirmed. Public benchmarks such as tau-bench show the pattern of an agent working against a simulated customer and a rule book. Real tickets supply the long tail of requests and the policies that a synthetic task set misses. More on that on the RL environments page.

De-identification specific to tickets

Tickets carry structured identifiers in free text: order numbers, invoice references, licence keys, addresses and bank details typed by customers who wanted a quick answer. Those must become tokens or be removed. Customer organisation names belong in the same treatment as people. Check satisfaction comments too, since customers often name the agent who helped them.