# MemCat banking classification example: provenance

The downloadable index is an example for 77 English banking support categories. It proposes labels, never performs financial actions. The SDK itself includes no model weights.

- BANKING77: Iñigo Casanueva, Tadas Temcinas, Daniela Gerz, Matthew Henderson and Ivan Vulić / PolyAI, *Efficient Intent Detection with Dual Sentence Encoders* (2020). [Paper](https://arxiv.org/abs/2003.04807), [source](https://github.com/PolyAI-LDN/task-specific-datasets/tree/57ec275d8078af65b7731c2a98be812d844a6d6b/banking_data), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Original labels preserved. Normalized same-label duplicates removed before a recorded training/development/calibration split. Up to 32 training examples per category plus its humanized label description were embedded; the index contains derived vectors. No official test rows were read or used to construct prototypes or select thresholds; seven normalized text overlaps inherited from the source split are disclosed in /evidence.
- Encoder: [Xenova/all-MiniLM-L6-v2](https://huggingface.co/Xenova/all-MiniLM-L6-v2/tree/751bff37182d3f1213fa05d7196b954e230abad9), revision `751bff37182d3f1213fa05d7196b954e230abad9`, Apache 2.0. Quantized q8 ONNX, mean pooling, L2 normalization, 384 dimensions, max 256 tokens. Runtime Transformers.js 4.3.0; browser WASM on one thread. Browser downloads public assets only after explicit loading.
- Separate distant-negative diagnostic: Stefan Larson et al., *An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction* (2019), [CLINC150 source](https://github.com/clinc/oos-eval/tree/828f8093932c8fe6ca7936c3d2e52903b1c523de), [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/). All 30 official test records from each of eight declared nonbanking intents: weather, recipe, translate, time, alarm, timer, play_music, restaurant_reservation. This subset cannot establish near-domain unknown performance.

Public datasets may have appeared in encoder pretraining; this is not a contamination-free claim. Authored stress examples are separate regression evidence. Cosine similarity is not a calibrated correctness probability. Full measurements, errors, hashes and caveats are linked from [/evidence](https://usememcat.vercel.app/evidence).

Dataset and model licenses do not grant a license to the separately distributed private MemCat SDK.