SMS Spam Collection — ChatGPT-Extended & De-Duplicated
Description
This dataset is an extended and de-duplicated version of the widely used UCI SMS Spam Collection. It combines the original SMS spam corpus with additional spam and ham messages generated using ChatGPT. The generated candidates were subjected to a multi-stage validation and verification process before being considered for inclusion. Exact duplicates, malformed records, missing values, and generated messages that did not satisfy the validation criteria were removed. Dataset Composition: The initial merged dataset contained 10,572 records consisting of the original corpus and generated candidate messages. After data cleaning, label verification, duplicate removal, and validation, 5,173 records were retained: Ham: 4,521 messages (87.4%) Spam: 652 messages (12.6%) The final dataset therefore remains strongly imbalanced toward the ham class, which is consistent with the distribution commonly encountered in SMS spam classification. The substantial reduction from 10,572 raw records to 5,173 validated records is intentional. Many generated candidates were removed because they were exact or near-duplicates of existing messages, did not satisfy the required label, or did not sufficiently meet the validation criteria. Consequently, this dataset should be regarded as a targeted synthetic extension rather than a large-scale synthetic replacement of the original corpus. The dataset should not be interpreted as a collection of independently verified real-world SMS messages because some records are synthetically generated. FILE FORMAT: CSV (UTF-8) Final validated dataset: Total rows: 5,173 Ham: 4,521 Spam: 652 Missing values: None in the distributed final CSV Duplicate messages: Removed from the final validated dataset If the raw merged file is distributed separately, it contains 10,572 records before the final cleaning and de-duplication process.
Files
Steps to reproduce
STEPS TO REPRODUCE Download the UCI SMS Spam Collection and load the original Category and Message fields. Generate additional ham and spam messages using a fixed ChatGPT prompt, while recording the target label, prompt, batch, and generation details. Merge the original and generated messages into a single working dataset. Verify generated labels using an independent classifier-consensus check and manual spot-checking. Remove exact duplicates and filter near-duplicates using sentence embeddings and a predefined cosine-similarity threshold. Clean the dataset by removing malformed labels, missing values, and invalid messages. Compare distributions such as message length, vocabulary overlap, and readability between authentic and generated messages to identify major synthetic artefacts. Verify the final dataset, obtaining 5,173 unique records: 4,521 ham and 652 spam. Evaluate downstream performance by comparing models trained on the original corpus and the extended dataset using the same train/test protocol and metrics. Preserve all preprocessing and generation settings—including prompts, model information, similarity threshold, validation rules, and random seeds—to ensure reproducibility.