Anonymized Multiplatform Indonesian Social Media Comments on the Free Nutritious Meals (MBG) Program: Stance Labels from GPT and Gemini

Published: 18 November 2026| Version 1 | DOI: 10.17632/5ytvhs4jzk.1
Contributors:
Daffa Tarigan,
,

Description

This dataset contains 72,225 anonymized comments from public Instagram, TikTok, YouTube, and X/Twitter posts discussing Indonesia’s Makan Bergizi Gratis (MBG) program. The corpus combines comments from selected posts and videos with posts retrieved through three MBG-related X/Twitter searches. The CSV includes comment text, platform and post context, timestamps and available engagement metadata. Direct identifiers and selected personal information in the text have been masked, record identifiers have been hashed, and TikTok tracking parameters have been removed. The dataset also includes stance and confidence labels from Gemini 4 Argon and GPT-6.1 Sol, matched by record_id_hash. A 500-item sample has separate fit, stance, and sarcasm annotations from two human raters per item, along with agreement fields. Human stance agreement is present for 351 fit items; sticker-only items absent from the corpus and stance disagreements are not included in that consensus subset.

Files

Steps to reproduce

Select the MBG-related public sources listed 9 Instagram posts, 14 TikTok videos, and 7 YouTube videos. Collect X/Twitter posts using the three recorded queries: "Makan Bergizi Gratis", ("makanan bergizi gratis" OR "mbg"), and "makanan bergizi gratis". Collect comments and available post, comment, timestamp, and engagement metadata for the selected sources. Bright Data was used for part of the collection workflow. The project registry records the targets and collection method categories, but not the Bright Data scraper names, run settings, or export files; retain those original materials with the project if available. Export the source data as six CSV files: Instagram, TikTok, YouTube, and three X/Twitter search exports. Consolidate them into one table while retaining platform and source_file fields. Load all record and post IDs as strings to preserve their exact values. Clean and anonymize the consolidated data. Remove empty comments and duplicate records; replace account handles, personal names, locations, email addresses, and phone numbers with anonymized placeholders; hash record IDs and parent-comment IDs; standardize timestamps to UTC; and remove TikTok tracking parameters from post URLs. Save the anonymized corpus as MBG-Anon.csv. The current release contains 72,225 records. Join the Gemini 4 Argon and GPT-6.1 Sol stance and confidence outputs from mbg-anon-gemini.csv and mbg-anon-gpt-61-sol.csv to MBG-Anon.csv using record_id_hash. Add the human annotations from the 500-item workbook sample. Preserve each annotator’s fit, stance, and sarcasm labels separately. Match sample comments to corpus records using platform, normalized comment text, and post context URL. Two sticker-only sample items are not present in the release; repeated identical comments can map one sample item to multiple corpus rows. For stance evaluation, compare model predictions with sample items where both human annotators marked the comment as fit and agreed on stance. Leave stance out of evaluation for not_fit items and human stance disagreements. In the current comparison, 351 sample items met this agreement rule.

Institutions

Categories

Social Sciences

Licence