Ineeoluwa/support-intent-grouped

Support Intent Grouped

240 synthetic English support requests, six labels, with deterministic, label-stratified splits that keep entire template families together. Created by Inioluwa Atanda for AdapterLab.

Split	Examples	Groups	Examples per label
Train	180	36	30
Validation	30	6	5
Test	30	6	5

Labels: account, billing, cancellation, delivery, login, technical. Each record contains text, label, group, and split. The label is a string; map labels in sorted order when training a classifier. All names, reference identifiers and requests are fictional. There are no customer records or calls.

Why group the split?

Each label has eight hand-specified template families and five variants per family. Variants share a family ID. A naive random row split would put nearly identical patterns in training and test. Here, SHA-256 ordering of 42:{group} assigns one family per label to test, one to validation and six to training. All 48 families appear in exactly one split. This reduces template leakage; it does not make a synthetic dataset representative of real users.

dataset_manifest.json records the canonical SHA-256 fingerprint: 452710ab0bf595050805180e10ff423b58723ae1f34c22153da411fd87cbd8dc. It matches the trained reference model's manifest. To reproduce, call prepare_dataset(demo_dataset()) from the source revision pinned in the model's provenance.json. The generator is in backend/adapterlab/data.py.

from datasets import load_dataset

dataset = load_dataset("Ineeoluwa/support-intent-grouped")
print(dataset["train"][0])

Uses and limitations

Use it for testing training products, grouped split logic and small classification experiments. It is not a natural-language benchmark or a customer-support corpus. Examples are short, single intent, English and highly templated. Reference tokens and suffixes may create shortcuts. There is no unknown-intent class, conversational context, spelling-noise study or multilingual coverage. The test set is only 30 examples and should not support production accuracy claims.

The reference LoRA model has test macro F1 0.5199 and includes all per-class failures. The data was generated for the real training run on 27 October 2026; source Git reconstruction dates do not date the dataset or experiment.

CC0-1.0 applies to these synthetic examples. No underlying customer data was used.