A 192-row intent dataset with a DOI is worth more than your bigger one without

Four test examples per class is where the arithmetic gets awkward ^[4]. Recall over four items can only land on 0, 25, 50, 75 or 100 percent, so a single error in one class moves that class's recall by a quarter ^[16]. Across the whole 32-row test split, one flipped prediction is worth 3.1 points of accuracy ^[15]. The macro F1 and confusion matrix the maintainers ask you to report ^[9] will cleanly separate a router that works from one that does not. They will tell you close to nothing about whether one embedding model beats another by two points. TekPedal says as much: this is a controlled regression suite, not evidence about customer behaviour ^[7].

The provenance is the part worth copying. Every row was written editorially from documented subject lists and templates rather than pulled from search logs, support tickets or customer conversations, and each row carries a `synthetic-editorial` marker so its origin travels with it downstream ^[6]. That is why the release can be CC BY 4.0 at all ^[2], and why the Zenodo DOI means something ^[3]: nobody redistributing it has to work out whether a given Turkish query started life in somebody's chat transcript.

This is precisely where most in-house intent data fails. Log-derived sets are better evidence and worse artefacts. They cannot be licensed outward, the splits move every time someone appends rows, and there is no frozen version an external reviewer can cite when your reported macro F1 goes up. The 192 rows here are fixed at 128 train, 32 validation and 32 test, with the allocation deterministic inside each class ^[4], and the repository runs automated checks on total row count, class balance, split sizes, ID uniqueness, required fields and common direct-identifier patterns ^[8]. That last check is the giveaway of someone who expects the data to be redistributed.

The bill for all this is stated openly by the maintainers: the corpus does not measure real search demand or the natural frequency of user intents ^[7]. Twenty-four rows per class ^[1] is a design decision. In production, a misroute on the most common class costs more than a misroute on the rarest, and this dataset deliberately contains no information about which is which.

One design detail earns its place. Each record carries not just the intent identifier and a Turkish label but a suggested TekPedal content route, plus the split, language and provenance marker ^[5]. The labels are wired to the product's own destinations rather than to abstract categories, which is what makes the set testable as a router rather than only as a classifier. The stated reason for splitting intents this way is that a station lookup needs a location index, a price question needs current tariff data, and a battery question needs educational content ^[12].

What is missing is the failure mode that actually costs money. Coverage of spelling errors, dialect, code-switching, multi-intent and out-of-scope queries is limited, and TekPedal says a real deployment still needs monitoring, rejection behaviour and broader evaluation data ^[10]. A router with no out-of-scope class in its test split has not been tested on the thing it will do worst: answering confidently when the question belongs to none of the eight ^[11].