sovereigndeveloper/mitw-kym-meme-interpretation

MITW-KYM: A Validated Multimodal Meme Interpretation Dataset

Dataset Summary

MITW-KYM is a small validated multimodal meme interpretation dataset containing 105 selected meme images. The dataset focuses on cases where meaning emerges through image-text interaction, pragmatic inference, cultural context, ambiguity, incongruity, or potential false-positive moderation risk. Each item was selected by a human researcher and validated using two frontier multimodal LLM annotators: GPT-5.5 and Gemini 2.5 Pro. Inclusion required 2/3 agreement between the human researcher and the two LLM annotators.

The dataset is intended for research on multimodal interpretation, harmful meme analysis, pragmatic ambiguity, and model disagreement. It is not intended as a general-purpose hate speech benchmark or a comprehensive sample of online hate.

---

Intended Use

• Research on multimodal reasoning in memes
• Study of pragmatic inference, ambiguity, and incongruity in image-text pairs
• Evaluation of model performance on culturally coded or contextually dependent content
• Analysis of human-LLM agreement on interpretive difficulty
• False-positive analysis for automated moderation systems

---

Data Sources

Images were collected from:

Source	Count
Know Your Meme (KYM)	~36
9GAG	~38
Gab	~18
4chan /pol/	~7
X / Twitter	~6

---

Annotation and Validation Procedure

Human Selection

A human researcher manually reviewed meme candidates from the above sources using four multimodal selection criteria (Q1–Q4) and assigned each accepted meme to one of five interpretive categories.

LLM Validation

The selected memes were independently evaluated by two frontier multimodal LLM annotators: OpenAI GPT-5.5 and Gemini 2.5 Pro.

For each meme, annotators judged four criteria:

• Q1 — Whether image-text interaction is essential (neither modality alone gives the full meaning)
• Q2 — Whether meaning requires inference beyond the literal surface
• Q3 — Whether cultural, contextual, or subcultural knowledge is required
• Q4 — Whether the meme contains incongruity, tension, reversal, or interpretive uncertainty

A meme was included in the final dataset if at least 2 out of 3 annotators — the human researcher, GPT-5.5, and Gemini 2.5 Pro — accepted it.

Results

Metric	Value
Total selected memes	105
Annotators	human researcher, GPT-5.5, Gemini 2.5 Pro
Required votes	3
Majority threshold	2/3
Passed majority	105
Failed majority	0

The human category label is retained as the canonical category. LLM labels are included as auxiliary validation metadata.

---

Category Schema

Each meme is assigned exactly one category:

Category	Description
explicit	Low-inference harmful or degrading content where the target or stance is relatively direct
ambiguous	Target, stance, or intended meaning is uncertain or underspecified
incongruous	The meme depends strongly on text-image tension, contradiction, reversal, or semantic mismatch
dogwhistle	Meaning depends on coded, subcultural, political, extremist, or otherwise contextual recognition
benign_misleading	Non-harmful or weakly harmful content likely to be falsely flagged by shallow moderation systems

Category Distribution

Category	Count
ambiguous	25
incongruous	24
explicit	20
dogwhistle	18
benign_misleading	18
Total	105

---

File Structure

hf_mitw_kym_dataset/
  README.md                          # This file
  LICENSE                            # CC BY 4.0 (annotations/metadata only)
  hf_dataset_final_105.jsonl         # Final merged dataset (105 rows)
  data/
    mitw_kym_selected.jsonl          # Human-selected candidates
    majority_report.jsonl            # Per-meme voting summary
    majority_passed.jsonl            # Memes that passed 2/3 threshold
    majority_failed.jsonl            # Memes that failed (empty)
    majority_incomplete.jsonl        # Memes with incomplete votes (empty)
  annotations/
    openai_gpt-5.5_annotations.jsonl # GPT-5.5 annotation records
    gemini_2.5-pro_annotations.jsonl # Gemini 2.5 Pro annotation records
  images/
    <candidate_id>.<ext>             # Meme images
  scripts/
    build_hf_dataset.py             # Builds hf_dataset_final_105.jsonl
    verify_dataset.py               # Integrity check

---

Data Fields

Each row in hf_dataset_final_105.jsonl:

Field	Type	Description
candidate_id	string	Unique identifier
image	string	Repo-relative image path (images/<id>.<ext>)
source	string	Origin platform
entry_title	string	Meme title (KYM entries) or filename
entry_url	string	Source URL (if available)
human_category	string	Canonical category assigned by human researcher
human_accepted	bool	Always true (all rows passed selection)
human_q_selections	list	Which Q criteria the human researcher selected
openai_gpt55_accepted	bool	GPT-5.5 inclusion decision
openai_gpt55_category	string	GPT-5.5 assigned category
openai_gpt55_gate_yes_count	int	Number of gate criteria GPT-5.5 answered YES
openai_gpt55_q1–q4	bool	GPT-5.5 per-criterion answers
gemini_25pro_accepted	bool	Gemini 2.5 Pro inclusion decision
gemini_25pro_category	string	Gemini 2.5 Pro assigned category
gemini_25pro_gate_yes_count	int	Number of gate criteria Gemini answered YES
gemini_25pro_q1–q4	bool	Gemini 2.5 Pro per-criterion answers
agreement_accept_count	int	Number of annotators who accepted (max 3)
agreement_required_votes	int	Total votes cast
agreement_threshold	int	Minimum accepts required for inclusion
included_in_final_dataset	bool	Always true
llm_majority_category	string	Category with most LLM votes
category_counts	object	Per-category vote counts
about	string	KYM entry description (where available)
origin	string	KYM origin section (where available)
spread	string	KYM spread section (where available)
text	string	OCR-extracted meme text (reserved, currently null)
original_local_image_path	string	Original local collection path

---

Limitations

• Small size: 105 memes is sufficient for qualitative analysis and model probing but not large-scale training.
• Single human annotator: Pass 0 uses a single researcher. Inter-annotator agreement on the full set has not yet been measured.
• English-language bias: All memes are primarily English-language.
• Source bias: KYM over-represents Western internet culture. The custom collection adds fringe platform content (4chan, Gab) which may not be representative.
• LLM validator limitations: GPT-5.5 and Gemini 2.5 Pro may share systematic biases. Disagreement between LLMs and the human researcher is a feature, not a bug — it is part of the dataset's analytical value.
• No OCR: The text field is currently null; meme text has not been extracted.

---

Ethical Considerations

This dataset contains content that may be harmful, offensive, or distressing, including political extremism, coded hate, and degrading imagery. It is released for research purposes only. Users should:

• Not use this dataset to train models intended for automated content generation
• Handle sensitive content in accordance with institutional ethics policies
• Not redistribute individual meme images outside of research contexts without assessing applicable rights

---

Licensing

Annotation metadata, validation records, and documentation: CC BY 4.0

Meme images: owned by their respective creators/platforms. Provided here for non-commercial academic research only. See LICENSE for full details.

---

Citation

@dataset{mitw_kym_2026,
  title        = {MITW-KYM: A Validated Multimodal Meme Interpretation Dataset},
  author       = {},
  year         = {2026},
  note         = {105 memes validated by human researcher, GPT-5.5, and Gemini 2.5 Pro}
}

(Citation details to be updated upon publication.)