---
license: cc-by-4.0
tags:
- concept-bottleneck-models
- phishing-detection
- nlp
---

# NWeak/emails-mirror

Preprocessed phishing-email concept-bottleneck data (see `emails/scripts/preprocessing.ipynb`
in [user-study-CBMs](https://github.com/debryu/user-study-CBMs)): sentence
embeddings (`all-MiniLM-L6-v2`), the full 19-dim concept ground truth
(`concept_gts`) and the 6-dim merged/filtered concept ground truth used for
modeling (`concept_gts_f` -- Fear+Authority combined, plus Urgency, Curiosity,
Neutral, Reply, Open attachment), and the task label (`label`: 0=Phishing,
1=Valid, 2=Spam).

`test` is `user_study_data.parquet` -- the 1000-email class-balanced subset
used for the concept-bottleneck user study (same split `train_concept_extractor2.ipynb`
evaluates against).

## Source dataset and attribution

The emails themselves are **not ours**. They come from the
**PhishingSpamDataSet** of Toth, Bisztray and Dubniczky, released under
CC-BY-4.0 and redistributed here under that licence:

> Rebeka Toth, Tamas Bisztray, Richard A. Dubniczky.
> *Constructing and Benchmarking: a Labeled Email Dataset for Text-Based
> Phishing and Spam Detection Framework.*
> [arXiv:2511.21448](https://arxiv.org/abs/2511.21448) |
> [github.com/DataPhish/PhishingSpamDataSet](https://github.com/DataPhish/PhishingSpamDataSet)

Theirs: `Subject`, `Body`, `Sender`, `URL(s)`, `Type`, `Created by`, `Source`,
`Year`, and the `LLM detected emotion` / `LLM detected motivation`
annotations. Ours: the filtering and split, the embeddings, the 6-concept
merge, and the trained models' outputs.

Please cite Toth et al. alongside our paper if you use this data.

**Note:** `Sender` and `Body` are real email content carried over from the
source dataset.

## Citation

Produced for **[Are Concept Bottleneck Models Effective as Decision-Support
Systems?](https://arxiv.org/abs/2608.25581)** (arXiv:2608.25581) -- Bogani,
Debole, Marconato, Pugnana, Tentori, Passerini. Code:
[github.com/debryu/user-study-CBMs](https://github.com/debryu/user-study-CBMs).