# CLINC150 shopping and payment intents

PRELIMINARY - modified Krummholz release, 2026-10-22.

Crowdsourced requests for order placement, order status and declined cards. The other intents and out-of-scope examples are excluded.

This is an intent-classification research corpus. It is not a production-ready support assistant or an authoritative source of policies or financial advice. The intended use does not add any restriction to the source license.

## Source, attribution and license

- Original source: https://github.com/clinc/oos-eval
- Provenance: Crowdsourced scenario and paraphrase corpus.
- Attribution: Clinc; Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang and Jason Mars. An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction (2019). https://aclanthology.org/D19-1131/
- License: CC-BY-3.0 - https://creativecommons.org/licenses/by/3.0/. See LICENSE.txt.
- Prepared from publisher files downloaded for this release. Source URLs and SHA-256 checksums are in source-manifest.json.
- Krummholz modified these data by scope filtering, field reduction, whitespace/Unicode normalization, automated privacy screening, template-slot standardization and deduplication. The original publishers have not endorsed this release.

## Scope and audit

The release preserves the latest published full 77-intent BANKING77 scope, which superseded the earlier ten-intent banking subset. CLINC150 and Bitext retain the agreed focused commerce/customer-service allowlists. FILTER-CONFIG.json contains every allowed source intent for this package. It uses all matching rows in the full source snapshot, without random subsampling.

- Source rows examined: 23,700.
- Out-of-scope rows excluded: 23,250.
- Rows in scope before cleaning: 450.
- Retained rows: 447, covering 3 intents.
- Rows with detected spans redacted before deduplication: 67.
- Normalized duplicate rows excluded: 3, including 3 cross-source duplicates and 0 same-source split collisions.
- Conflicting-label rows excluded: 0.
- Rows with an unresolved Unicode replacement character excluded: 0.
- Retained split counts: {"test": 89, "train": 298, "validation": 60}.

Detailed counts are in audit.json and intent-counts.csv. Exclusion counts are sequential; cross-source and cross-split counts are subsets of duplicate exclusions, not additional exclusions.

## Cleaning and privacy

Only the source utterance, intent and original split are retained as training content. Direct email, URL, IP-address, phone/long-number, identifier, social-handle, street-address and explicit-name patterns are replaced with placeholders. A pet name identified during source inspection is also redacted. spaCy en_core_web_sm 3.8.0 redacts detected person, location, organization, nationality/group, date, time and monetary entities. This deliberately generalizes some nonsensitive public brand names and relative dates as well. Existing synthetic template slots are standardized, never populated with identities. Raw responses, source row numbers, source flags and other unnecessary fields are excluded. Rows containing an unresolved Unicode replacement character are excluded rather than guessing the missing character.

Bitext response targets are omitted from this release because generic generated support policies and refund promises are not verified for any business. Source spelling variations are retained to support intent recognition.

Deduplication compares normalized text after privacy processing across all three customer-service packages. The key is Unicode NFKC plus case folding with punctuation and whitespace removed. Test is preferred, then validation, then train; source ties prefer BANKING77, then CLINC150, then Bitext. Contradictory labels for the same normalized text within one source are removed. This prevents exact/normalized overlap across published splits and these three packages. It does not detect all semantic paraphrases.

## Data dictionary

The CSV and JSONL contain the same records in UTF-8. Use either format, not both as separate examples. review-sample.csv is a small subset for inspection; do not concatenate it into training data.

| Field | Meaning |
|---|---|
| record_id | A content-derived identifier of the cleaned record; not a person or source account ID. |
| text | Cleaned user query. Bracketed uppercase values are redaction or template placeholders. |
| intent | Unchanged source intent label. |
| split | Original source train, validation or test designation. Bitext supplies a train-only corpus. |
| dataset_notice | Required release/modification/license notice. Exclude this field from model input features. |

Do not train on validation/test rows when evaluating. The retained evaluation subsets have been modified and are not directly comparable to original benchmark results. Bitext has no independent evaluation split; group semantic templates before constructing one.

## Limitations

PRELIMINARY means the cleaning is complete to the documented automated checks, while human label review, exhaustive personal-information review, semantic near-duplicate review and deployment validation remain outstanding. Automated entity detection can miss names or redact ordinary words. No claim of complete anonymization is made. The data are English and domain-limited, and the synthetic corpus does not represent real customer prevalence. The privacy transformation changes lexical and entity distributions. Test on representative, appropriately governed data before operational use.