diegovelilla/email-classification-dataset

Dataset Card for Email Classification Dataset

The Email Classification Dataset contains 10,900 English emails originally published as a binary spam/not_spam dataset and extended into a multiclass text classification task as part of the MLOps course at FIB-UPC (Master's in Data Science). Emails labeled not_spam in the original dataset were automatically re-labeled into multiple categories using an LLM.

Dataset Details

Dataset Description

This dataset contains 10,900 English emails for text classification tasks. Originally a binary spam detection dataset (spam/not_spam), the not_spam emails were automatically re-labeled into multiple categories using a Large Language Model (LLM), transforming the task into multiclass text classification. The dataset is split into train (8,175 rows, 75%) and test (2,725 rows, 25%) splits. It contains three columns: text (full email body), label (multiclass label assigned via LLM), and original_label (original binary label for traceability).

• Curated by: Carles Aguilera, Alex Bueno, Joel Delgado, Berta Torrents, Diego Velilla (MLOps Course, FIB-UPC)
• Language(s) (NLP): English
• License: Apache-2.0

Dataset Sources

• Repository: https://huggingface.co/datasets/Deysi/spam-detection-dataset

Uses

Direct Use

• Multiclass Email Classification: Train models to classify emails into multiple categories beyond binary spam detection.
• Binary Spam/Ham Email Classification: Use the original_label column for traditional spam detection tasks.
• Raw Text for LLM Training: The text column provides raw English email content suitable for language model pre-training or fine-tuning.

Out-of-Scope Use

• Models trained on this dataset should not be expected to generalize to real-world phishing tactics, as the spam text contains exaggerated, promotional language that may not reflect actual phishing attempts.
• Performance on non-English emails is not tested and likely to be worse.
• Not suitable for applications requiring expert-level multiclass annotation quality, as labels were LLM-generated and may contain noise compared to human expert labeling.

Dataset Structure

Dataset Instances

Each instance contains the following fields:

Field	Type	Description
text	string	Full email body as plain text
label	string	Final multiclass label assigned via LLM
original_label	string	Original binary label: spam or not_spam

Dataset Splits

Split	Rows	%
train	8,175	75%
test	2,725	25%
Total	10,900	100%

Dataset Creation

Curation Rationale

The dataset was originally published as a binary spam detection dataset. As part of the MLOPS course in the Master's in Data Science at FIB-UPC, the dataset was extended to support multiclass text classification by automatically re-labeling not_spam emails into multiple categories using an LLM. This transformation enables exploration of more nuanced classification tasks beyond binary spam detection.

Source Data

Data Collection and Processing

The original dataset was sourced from Deysi/spam-detection-dataset on Hugging Face, containing 10,900 English emails with binary spam/not_spam labels. The not_spam emails were then processed using an LLM to assign multiclass category labels. The original_label column was retained for traceability, preserving the original binary classification.

Who are the source data producers?

The original binary data was sourced from Deysi and published on Hugging Face. Information from the original dataset's sources is not clear.

Annotations

Annotation process

The multiclass labels derived from not_spam emails were generated automatically using a Large Language Model (LLM). The LLM was prompted to assign category labels to emails originally labeled as not_spam.

EXPLAIN HOW DID WE GENERATE THE MULTICLASS LABELS USING THE LLM

The original_label column is preserved for traceability and validation. The annotation process for the original binary labels is unkwnown.

Bias, Risks, and Limitations

• Synthetic spam style. Original spam texts contain exaggerated, promotional language that may not reflect real-world phishing tactics. Models trained on this dataset may not generalize to real-life spam.
• Temporal drift. Dataset likely reflects spam patterns from its original (and unknown) collection period. Current spam tactics may differ.
• English-only. No multilingual coverage. Performance on non-English email is not tested and likely to be worse.
• LLM annotation noise. The multiclass labels for not_spam emails have been LLM-generated and may differ from expert annotations.

Recommendations

Users should be made aware of the risks, biases, and limitations of the dataset. Specifically, users should:

• Be cautious when applying models trained on this dataset to real-world spam detection, as the spam style may not generalize.
• Consider validating multiclass labels with human experts if high annotation quality is required.
• Be aware that temporal and linguistic biases may limit applicability to other contexts.

Citation [optional]

BibTeX:

@dataset{email_classification_2026,
  author = {Carles Aguilera, Alex Bueno, Joel Delgado, Berta Torrents, Diego Velilla},
  title = {Email Classification Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Deysi/spam-detection-dataset}
}

APA:

Aguilera, C., Bueno, A., Delgado, J., Torrents, B., & Velilla, D. (2026). Email Classification Dataset. Hugging Face. https://huggingface.co/datasets/Deysi/spam-detection-dataset

Dataset Card Contact

This project has been carried out by:

• Carles Aguilera
• Alex Bueno
• Joel Delgado
• Berta Torrents
• Diego Velilla