sharifmmm/Sharif-EmojiVerse-72M

Silver-Label Multilingual Sentiment Distillation Data

Derived artifacts from the study "From 10K Labels to 72M Classifications: Scaling LLM Silver-Label Distillation for Multilingual Sentiment" (Ullah, HHAI-KEML 2026, CEUR-WS proceedings).

Preprint: https://zenodo.org/records/21786160

Dataset Summary

This repository contains derived artifacts from a study of LLM silver-label distillation on 72 million multilingual social media comments. It does not contain the raw comment corpus itself. What is released here is intended to support reproducibility of the paper's methodological claims and to enable downstream use of the distilled labels and aggregate statistics.

The full raw corpus is not publicly released due to social media platform terms of service and user privacy considerations. Access to the full corpus can be requested from the author for academic research use, subject to a data use agreement.

What This Dataset Contains

This dataset release includes:

• —Silver-label subsets: GPT-4o-mini generated silver labels (love/hate) for the 15,000 comments used in the study's silver-label experiments (sample IDs only, not comment text)
• —Held-out test set predictions: silver-label predictions on the 2,000-sample evaluation set used in the paper (sample IDs only)
• —Aggregate statistics: per-language sample distribution, emoji frequency counts on the sampled subset, comment length histograms, class balance statistics
• —Reproducibility subset (if included): a small hand-reviewed subset of generic, non-user-identifying comment content for methodological reproducibility

What This Dataset Does NOT Contain

• —The full 72M-comment corpus (not publicly redistributed)
• —Individual comment text with user identifiers
• —Usernames, timestamps, or any personally identifying information
• —Complete Instagram or TikTok post content

Intended Use

This dataset is intended for:

• —Reproducing the scaling and distillation analyses from the HHAI-KEML 2026 paper
• —Studying silver-label distributions and class imbalance patterns
• —Downstream research on multilingual data-efficient learning

This dataset is not intended for:

• —Individual user profiling
• —Content moderation training that requires raw text
• —Re-identifying original social media users
• —Commercial use of derived labels for a specific platform without that platform's authorization

Dataset Structure

<pre> ├── silverlabels/ │ ├── train15klabels.csv │ └── test2klabels.csv ├── aggregatestatistics/ │ ├── languagedistribution.csv │ └── classbalance.json └── README.md </pre>

Data Fields

silver_labels/train_15k_labels.csv:

• —sample_id: integer sample identifier (does not link to any external Instagram/TikTok post)
• —silver_label: string, either "love" or "hate"
• —source_language: ISO 639-1 language code (or "emoji-only" for emoji-only samples)

aggregate_statistics/language_distribution.csv:

• —language_code: ISO 639-1 code or "emoji-only"
• —count: number of samples in the stratified 50K sample
• —percentage: percentage of the stratified sample

Data Provenance

The underlying corpus consists of 72,327,992 comments collected from public Instagram and TikTok posts across 42 detected languages.

The corpus was collected for research purposes as part of an applied AI project . Silver labels were generated using OpenAI's GPT-4o-mini via the Batch API.

Ethical considerations:

• —Only publicly visible comments on public posts were included in the underlying corpus
• —No private accounts, direct messages, or account-restricted content was accessed
• —The raw corpus is not redistributed to protect user privacy

Limitations

• —Silver labels, not gold labels: All labels in this dataset are LLM-generated. They may inherit biases or errors from the teacher model (GPT-4o-mini). Downstream users should not treat these labels as ground truth for absolute sentiment.
• —Binary sentiment: The love/hate binary framing collapses nuanced emotional expression (sarcasm, mixed emotion, neutrality) into two classes.
• —Class imbalance: The label distribution is 95.4%/4.6% love/hate, reflecting the typical skew of public social media comments.
• —Language detection errors: Language labels were assigned via langdetect on short texts and may be systematically wrong for very short comments. See the paper's Limitations section for details.
• —Latin-script bias in evaluation: Per-language analysis in the paper covers only Latin-script languages that reached n≥30 in the test set. Non-Latin scripts (Arabic, Cyrillic, Devanagari, CJK) are underrepresented in per-language analysis.

Citation

If you use this dataset or the accompanying model in your research, please cite:

@inproceedings{ullah2026silverlabels,
 title = {From 10K Labels to 72M Classifications: Scaling LLM Silver-Label Distillation for Multilingual Sentiment},
 author = {Ullah, Sharif},
 booktitle = {HHAI-KEML 2026: 2nd International Workshop on Knowledge Engineering meets Machine Learning, co-located with HHAI 2026},
 series = {CEUR Workshop Proceedings},
 publisher = {CEUR-WS.org},
 year = {2026},
 url = {https://zenodo.org/records/21786160}
}

License

The derived artifacts in this repository are released under CC BY 4.0.

The underlying raw comments (not included here) remain the intellectual property of their original authors and Instagram / TikTok. No claim is made over the raw content.

Contact

Sharif Ullah — md.sharif.ullah.forhad@gmail.com,sharif@bhbfc.gov.bd Personal page: https://forhadsidhu.github.io/sharif/

Acknowledgments

Thanks to the two anonymous HHAI-KEML 2026 reviewers whose feedback strengthened this work, and to collaborators at the University of New Orleans.

Related Resources

• —Paper preprint: https://zenodo.org/records/21786160