notd5a/sms-malicious-benign-dataset

Smishing & Spam Detection Dataset

A curated, English-only dataset for training binary text classifiers to detect smishing (SMS phishing) and spam messages. This dataset has been iteratively refined through multiple model versions, with each round of error analysis feeding back into data quality improvements.

---

Overview

Property	Value
Task	Binary classification (spam/smishing detection)
Language	English only
Full dataset size	~700k rows
Cleaned + diverse dataset	~196k rows (v0.2.2)
Label format	0 = benign, 1 = spam/smishing
Engineered features	23 columns
File format	CSV

---

Files

File	Rows	Description
augmented_dataset_enriched.csv	697,971	Full English-filtered dataset (raw)
augmented_dataset_enriched_undersampled.csv	126,628	3:1 undersampled dataset (v0.1)
dataset_final.csv	~196k	5:1 undersampled, cleaned, enriched with diverse samples + 23 features (v0.2.2, recommended)

---

Dataset evolution

This dataset has gone through three iterations, each driven by error analysis on the trained model:

v0.1 — Initial release

• 697k raw messages undersampled to 126k at 3:1 ratio
• 15 engineered features
• Trained the v0.1 DeBERTa model (F1: 0.93, AUC: 0.99)

v0.2 / v0.2.1 — Rebalanced + label audit

Error analysis on v0.1 revealed 341 false positives and 283 false negatives with systematic patterns. Key findings:

• 7 mislabeled examples found — phishing messages (SBI YONO, NHS COVID, Apple Pay, AnPost customs) incorrectly labeled as benign. These were corrected via a label audit.
• Undersampling ratio changed to 5:1 (from 3:1) — the 3:1 ratio was overrepresenting spam relative to real-world distribution, and the model had enough capacity to handle 5:1 with focal loss + class weights.
• 8 new engineered features added (see feature table below) targeting obfuscation evasion and short message patterns.

v0.2.2 — Cleaned + diversified (current, recommended)

Error analysis on v0.2.1 revealed that ~288 false negatives were truncated/tiny messages, and ~408 false positives came from benign categories underrepresented in training data. Changes:

• Dataset cleanup pipeline applied:

  • Removed truncated messages (cut mid-sentence from CSV parsing/scraping issues)
  • Removed encoding-corrupted messages (mojibake, control characters)
  • Removed exact and near-duplicates
  • Removed unusable fragments (<15 chars or <3 words)
  • Removed garbage content (pure numbers, email headers, markup)

• Diverse benign samples added from:

  • NUS SMS Corpus — 67k conversational SMS messages (real person-to-person texts)
  • UCI SMS Spam Collection — 4,827 ham messages
  • Mendeley SMS Phishing Dataset — 4,844 ham messages
  • Synthetic hard negatives — ~5k LLM-generated messages in 7 categories directly targeting false positive failure modes: retail promotions, bank transaction alerts, telecom notifications, delivery updates, appointment reminders, OTP/verification codes, and casual conversation with spam-like keywords

---

Label distribution

Full dataset (raw)

Label	Count	Share
0 (benign)	666,314	95.5%
1 (spam/smishing)	31,657	4.5%

v0.2.2 cleaned + diverse (recommended)

Label	Approximate count	Ratio
0 (benign)	~163k	5:1
1 (spam/smishing)	~33k	—

---

Dataset composition

This dataset was compiled and cleaned from 4 primary source datasets plus diverse supplementary sources:

Primary sources

#	Source	Key columns used
1	Phishing message dataset — Kaggle	text, label
2	SMS Phishing Dataset for Machine Learning — DOI: 10.17632/f45bkkt8pr.1	TEXT, LABEL
3	shaghayegh-hp/Smishing_Dataset	message, spam label, smishing label
4	www.smishtank.com	MainText, Malicious, Phishing, Suspicious, Malware

Supplementary sources (v0.2.2)

Source	Type	Messages	Purpose
NUS SMS Corpus	Public dataset	~67k benign	Conversational diversity
UCI SMS Spam Collection	Public dataset	~4.8k benign	Classic benchmark ham messages
Mendeley SMS Phishing Dataset	Public dataset	~4.8k benign	Different distribution ham messages
Synthetic hard negatives	LLM-generated	~5k benign	Targets FP failure modes (promos, bank alerts, OTPs)

Label normalisation

All source labels were mapped to a unified binary scheme:

• Source 1: All entries were phishing → mapped to 1
• Source 2: ham → 0, spam / smishing (any case) → 1
• Source 3: spam label == 1 OR smishing label == 1 → 1, otherwise 0
• Source 4: Any of Malicious, Phishing, Suspicious, or Malware > 0 → 1, otherwise 0
• Supplementary sources: All benign → 0
• Label audit corrections: 7 high-confidence mislabeled phishing messages flipped 0 → 1

---

Preprocessing

Original pipeline (v0.1)

1. Column standardisation — all sources normalised to message, label columns
2. Deduplication — duplicate rows removed on the message column
3. Language filtering — non-English rows removed
4. Undersampling — majority class randomly downsampled (seed=42)

Cleanup pipeline (v0.2.2)

The following additional steps were applied to produce the recommended dataset:

5. Truncation removal — messages cut mid-sentence (from CSV parsing, scraping, or encoding issues) detected and removed
6. Encoding repair — mojibake patterns (UTF-8 decoded as Latin-1), control characters, and replacement characters removed
7. Fragment removal — messages under 15 characters or under 3 words removed as unusable
8. Garbage removal — pure numbers, email headers, single repeated characters, code/markup
9. Near-duplicate removal — messages normalised (lowercase, URLs/numbers replaced) and hashed; clusters deduplicated keeping the longest variant
10. Label audit — 7 high-confidence mislabeled examples corrected (phishing → spam)
11. Diverse benign injection — public dataset ham messages + synthetic hard negatives added
12. Feature enrichment — all 23 engineered features computed and saved to CSV

---

Engineered features

The dataset includes 23 pre-computed features extracted from the message text. These are designed to complement DeBERTa-v3's semantic understanding with explicit structural and pattern-based signals.

Original 15 features (v0.1)

#	Feature	Type	Description
1	char_count	int	Total number of characters
2	word_count	int	Total number of words
3	avg_word_length	float	Average number of characters per word
4	uppercase_ratio	float	Proportion of letters that are uppercase
5	digit_ratio	float	Proportion of characters that are digits
6	special_char_ratio	float	Proportion of non-alphanumeric, non-space characters
7	exclamation_count	int	Number of ! characters
8	question_mark_count	int	Number of ? characters
9	has_url	binary	1 if any URL is present
10	url_count	int	Number of URLs detected
11	has_shortened_url	binary	1 if a known URL shortener is detected
12	has_phone_number	binary	1 if a phone number pattern is detected (≥7 digits)
13	has_email	binary	1 if an email address is detected
14	has_currency	binary	1 if a currency symbol or code is present
15	urgency_score	int	Count of urgency-related keywords matched

New 8 features (v0.2)

These features were added to target specific failure modes found in v0.1 error analysis — particularly obfuscated spam that evaded the original 15 features.

#	Feature	Type	Targets	Description
16	unicode_ratio	float	Unicode evasion ("Vérífy yøur àccount")	Non-ASCII characters / total characters
17	char_entropy	float	Template spam, repetitive content	Shannon entropy over character distribution
18	suspicious_spacing	int	Spaced-out evasion ("m e s s a g e")	Count of space-separated character sequences
19	leet_ratio	float	L33t speak substitutions	Characters that map to leet translations / total
20	max_digit_run	int	Embedded phone/account numbers	Longest consecutive digit sequence
21	repeated_char_ratio	float	Exclamation/character spam	Consecutive repeated chars / (length − 1)
22	vocab_richness	float	Low-diversity template spam	Unique words / total words
23	has_obfuscated_url	binary	Broken/evasive URLs ("httpscluesjdko")	Regex detection of obfuscated URL patterns

Model performance on this dataset

Models trained on successive versions of this dataset show the impact of data quality improvements:

Model	Dataset version	F1	Precision	Recall	AUC-ROC
v0.1 (DeBERTa-v3-base)	3:1 undersampled, 126k	0.9299	0.8986	0.9634	0.9906
v0.2.1 (DeBERTa-v3-base + v0.2 arch)	5:1 undersampled, ~196k	0.9022	0.9058	0.8986	0.9857
v0.2.2 (DeBERTa-v3-base + v0.2 arch)	6.3:1 cleaned + diverse	0.9096	0.9170	0.9023	0.9883

The v0.2.2 dataset achieves the highest precision (0.917) of any version — fewer legitimate messages incorrectly blocked — while maintaining strong recall and AUC.

---

Known issues and limitations

• English only. Non-English messages were filtered out. The dataset is not suitable for multilingual spam detection without additional data.
• Temporal bias. Source datasets span different time periods. Recent smishing campaigns (2025+) may use patterns not represented here.
• Synthetic data. The v0.2.2 dataset includes ~5k LLM-generated benign messages. While designed to be realistic, synthetic data may not fully capture real-world message diversity.
• Label noise. A label audit corrected 7 high-confidence errors, but medium and low-confidence label issues may remain. The full audit flagged ~7.4k rows for manual review.
• Promotional boundary. Legitimate marketing messages with aggressive urgency language remain the hardest category to distinguish from spam. This is a fundamental challenge in the domain, not a dataset deficiency.

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You are free to use and share this dataset for non-commercial and research purposes with appropriate attribution. Commercial use is not permitted without explicit written permission from the author.

See https://creativecommons.org/licenses/by-nc/4.0/ for full license details.

Note: This dataset is compiled from multiple source datasets. Users are responsible for ensuring their intended use complies with the licenses of the underlying sources.

---

Disclaimer

This dataset contains real-world spam and phishing messages collected for research purposes. It is intended solely for the development of defensive security tools and classifiers. Misuse of this data is strictly prohibited.