somosnlp-hackathon-2025/memes_instagram_chilenos_es_small

memesinstagramchilenosessmall

A dataset designed to train and evaluate vision-language models on Chilean meme understanding, with a strong focus on cultural context and local humor, built for the Somos NLP Hackathon 2025.

<img src="./tralalelo-tralala-logo.png" alt="Banner" width="50%" />

Introduction

Memes are rich cultural artifacts that encapsulate humor, identity, and social commentary in visual formats. Yet, most existing datasets focus on English-language content or generic humor detection, leaving culturally grounded interpretations underexplored—especially in Latin America.

memes_instagram_chilenos_es_small is the first dataset that systematically aligns Chilean meme content with instruction-based, culturally contextual explanations in Spanish. It is designed to help models not only "describe" images but to understand why something is funny or meaningful in a Chilean context.

Motivation

Despite the significant advancements in vision-language models, understanding memes remains a complex task due to their ambiguity, irony, and strong cultural grounding. While many benchmarks exist for general image-text tasks, there were no public datasets offering culturally contextualized explanations for Chilean memes—until now.

The only known prior work, Chilean Memes (Poblete et al., 2020), introduced a dataset for classification and information retrieval, but it did not address the interpretability or cultural semantics behind the memes.

This gap motivated the creation of memes_instagram_chilenos_es_small, a dataset designed to:

• —Provide high-quality, culturally aligned explanations for memes popular in Chile.
• —Enable multimodal instruction tuning focused on local humor.
• —Support research into Latin American visual language, humor detection, and sociocultural reasoning.

Impact

This dataset contributes to the field in several ways:

• —It enables fine-tuning and evaluation of models in Spanish, with a regional and cultural focus.
• —It serves as a benchmark for multimodal cultural understanding, especially in the Latin American context.
• —Its pipeline—scraping, contextual enrichment, manual curation, and instruction templating—can be replicated for other regions or cultures.

The dataset has already been used in the development of `llava-v1.6-mistral-7b-memes-chilenos-small`, a fine-tuned vision-language model adapted to Chilean meme humor.

Dataset Summary

This dataset contains 4,776 examples, derived from 1,194 unique memes collected from popular Chilean Instagram pages. Each meme was enriched with culturally grounded explanations.

The dataset was created through the following pipeline:

1. 1.Scraping: Using Apify to extract memes and metadata from Instagram accounts.
2. 2.Text generation: GPT-4o was used to generate high-quality contextual descriptions based on post comments and captions.
3. 3.Manual annotation: All outputs were curated, refined, and validated to ensure cultural fidelity.
4. 4.Instruction tuning: Each meme received 4 different instruction-response pairs for diversity.

Features

Each example includes:

• —image: the meme image (PIL format).
• —url: the original Instagram post URL.
• —instruction: a culturally specific prompt.
• —multiple_choice_answer: a colloquial explanation aligned with the instruction.

Limitations

• —Demographic bias: The data skews toward urban, millennial/Gen Z Instagram audiences.
• —Temporal sensitivity: Meme relevance may decay over time.
• —Subjectivity: Humor interpretation is subjective and culturally dependent.

Citation

BibTeX

@dataset{memes_instagram_chilenos_2025,
  author       = {Somos NLP Hackathon 2025},
  title        = {memes_instagram_chilenos_es_small},
  year         = 2025,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/somosnlp-hackathon-2025/memes_instagram_chilenos_es_small}
}

Hackathon: This model was developed for Somos NLP Hackathon 2025 – see the project page here.

Resources

• —🔗 GitHub repository – Source code, training scripts, and the Gradio demo.