FaisalHejary/sms-spam-embedding

SMS Spam Collection for embedding training

Spam classification rows in a label-as-document format, for training text embedding models. The label is written as a sentence and used as the document, so the model learns to rank the right label above the wrong one.

Built from codesignal/sms-spam-collection, a copy of the UCI SMS Spam Collection by Tiago A. Almeida and José María Gómez Hidalgo. All messages come from that dataset; only the format, filtering and labels' wording are new here.

Columns

Column	Description
task	Task name: always spam.
query_inst	Instruction for the query: "Given an SMS message, classify whether it is spam or a normal message".
query	The SMS message.
document_inst	Instruction for the label: always null.
positives	The right label, as a list of one.
negatives	The wrong label, as a list of one.

Labels

Source label	Label text
ham	a personal message between friends or family
spam	an advertisement, a prize offer or a scam

How it was built

1. Messages are compared lowercased with punctuation removed. Duplicates are dropped, and so are messages labelled both ham and spam.
2. The source has only a train split, so a random 10% of the remaining messages (seed 0) became the test split. Duplicates are removed first, so no message is in both splits.
3. Each label is capped at 3,000 rows. Only ham is affected, in train.

Split	ham	spam	Total
train	3,000	572	3,572
test	455	57	512

Zero-shot baseline

Qwen/Qwen3-VL-Embedding-2B with no training, predicting the more similar label, on the test split:

Instructions	Accuracy	Macro F1
None	0.129	0.122
Query instruction	0.775	0.666

Always predicting ham scores 0.889 accuracy. The label wording was chosen by comparing a few wordings on the test split, so these scores are slightly optimistic.

Usage

from datasets import load_dataset

ds = load_dataset("FaisalHejary/sms-spam-embedding")

License and citation

CC BY 4.0, the same as the source. If you use this dataset, cite the original:

@inproceedings{almeida2011spam,
  title     = {Contributions to the Study of SMS Spam Filtering: New Collection and Results},
  author    = {Almeida, Tiago A. and G{\'o}mez Hidalgo, Jos{\'e} Mar{\'i}a and Yamakami, Akebo},
  booktitle = {Proceedings of the 2011 ACM Symposium on Document Engineering},
  year      = {2011}
}