Zovda
Support teams produce a stream that most machine learning teams would pay for: a user describing a problem in their own words, an agent resolving it, and a record of whether the resolution held.
It is usually archived, retention-limited, and never read as a dataset.
What makes a ticket valuable
- The question is real traffic, not a constructed example
- It carries the missing context the user forgot, added during the exchange
- The resolution was verified by the outcome, not by a rating
- The frustration is visible, which marks the difficult cases
The third item is what a benchmark rarely has: a check.
Clean before you learn
Tickets contain personal data, internal references, and the habits of one support team. Strip the first, map the second, and treat the third as a bias to name.
A model trained on raw tickets learns the team's shorthand along with the answer.
Split by problem, not by date
Splitting randomly leaks: the same problem appears in both halves with slightly different wording. Cluster by problem, then split by cluster, and the evaluation stops being a memorization test.
Keep the resolution type
Fixed by a setting, fixed by a change, answered with a policy, or escalated. Four types, four different training uses.
Zovda builds these collections through Datuva, where each ticket keeps its resolution type and its verification. The question bank collection shows a similar structure.
Tell us what your tickets already contain. We reply within one business day.