---
license: other
license_name: ibm-tabformer-derived
pretty_name: cortex embeddings for IBM TabFormer fraud benchmark
tags:
  - tabular
  - fraud-detection
  - embeddings
  - transaction-data
size_categories:
  - 10M<n<100M
---

# cortex embeddings — IBM TabFormer

Per-transaction hidden states (128-d, float32) produced by Neospace's **cortex** transaction
foundation model over every transaction in the public IBM TabFormer credit-card dataset
(`card_transaction.v1.csv`, 24,386,900 transactions, 2,000 cardholders, 1991–2020).

These embeddings let you reproduce the [NeoLDM benchmark](https://github.com/neospace-ai/NeoLDM)
without running cortex: feed them to the gradient-boosted-tree fraud classifier in that repo.

## Contents

| dir | checkpoint | rows |
|---|---|---|
| `pretrain/` | self-supervised (full-MLM; never saw `is_fraud`) | 24,386,900 |
| `finetune/` | supervised (finetuned on `is_fraud`) | 24,386,900 |

Each parquet has three columns:

- `user` — cardholder id (matches the IBM TabFormer `User` field).
- `timestamp__orig` — transaction timestamp (UTC, minute resolution).
- `hidden` — the 128-d cortex hidden state (`list<float>[128]`, float32).

Join key to the raw transactions: `(user, timestamp__orig)`. The encoding is `byte_stream_split`
+ zstd; precision is unchanged (float32), so results reproduce bit-for-bit.

## Usage

```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="<your-hf-org>/cortex-ibm-tabformer-embeddings", repo_type="dataset",
                  local_dir="artifacts/embeddings", allow_patterns=["pretrain/*", "finetune/*"])
```

Then point the NeoLDM scripts at `artifacts/embeddings/pretrain` (nb02) and
`artifacts/embeddings/finetune` (nb03). See the repo's README for the full pipeline.

## License & attribution

The embeddings are derived from the IBM TabFormer dataset (Padhi et al., *Tabular Transformers for
Modeling Multivariate Time Series*, [arXiv:2011.01843](https://arxiv.org/abs/2011.01843),
[github.com/IBM/TabFormer](https://github.com/IBM/TabFormer)) and are subject to its terms. They are
learned representations, not the raw transaction fields. The NeoLDM code is Apache-2.0.