kareudon/kaomoji-tagged

Kaomoji Labeled Dataset

At a glance:

• Rows: 47,311
• Format: parquet and jsonl
• Main unit: one kaomoji per row
• Core fields: id, kaomoji, tags, primary_tags, secondary_tags, source, quality_score, llm_confidence
• No train/validation/test split is included. This dataset was assembled as a reusable resource, not as a benchmark.

What this dataset is

This is a cleaned, tagged kaomoji dataset built on top of the public GitHub repository:

• Source repository: https://github.com/ekohrt/emoticon_kaomoji_dataset

The original repository provides a large kaomoji collection with two tag fields per item: original_tags and new_tags. This project takes that raw material, runs it through a rule-based cleaning pipeline, uses an LLM to fill in missing or weak labels, reviews newly proposed tags by hand, and then exports a stable dataset that other people can browse, filter, and reuse.

The end result is not a verbatim mirror of the source repository. It is a derived dataset with normalized tags, deduplicated samples, low-quality or broken records removed, and approved expanded tags merged into the public output.

Files in this folder

• data.parquet: the main release file and the one you will usually want on Hugging Face
• data.jsonl: the same rows in line-delimited JSON
• schema.json: field-level schema
• README.md: this dataset card

Data schema

Each row has the following fields.

Field	Type	Meaning
id	string	Stable sample identifier in this export
kaomoji	string	The kaomoji itself
tags	list[string]	Final deduplicated tag list
primary_tags	list[string]	Main semantic tags predicted for this kaomoji
secondary_tags	list[string]	Supporting tags; approved proposed tags are merged here
source	string	Provenance marker for this release
quality_score	float64	Rule-based quality score from the cleaning pipeline
llm_confidence	float64	Confidence score returned by the LLM annotator

Notes:

• tags is built from primary_tags + secondary_tags, with duplicates removed while preserving order.
• This public export does not expose style_tags as a separate field.

Dataset composition

The released dataset has 47,311 retained rows.

The broader processing run looked like this:

• 62,149 records were loaded from the original source dataset at the start of the pipeline.
• 48,323 records entered the final strict LLM annotation stage.
• 1,012 records were later dropped because the model marked them as reject=true.
• 47,311 records were kept in the final public export.

The annotation process also used a small manually labeled supplement of high-frequency personal-use kaomoji. Those samples were passed through the same later stages so they would not become a second-class side dataset.

Tag system

This project uses two layers of tags.

1. Project-defined base tags

The rule pipeline starts from a fixed canonical inventory of 58 tags. These are the tags the project explicitly set out to support.

They roughly fall into these groups:

• emotions and states
• facial actions and expressions
• visual or body-part cues
• a small style-oriented subset that existed in the project inventory

Examples include happy, sad, angry, love, crying, smiling, waving, blushing, heart, tears, hands, and eyes.

2. Approved expanded tags

During the LLM stage, the model was allowed to suggest one proposed_new_tag when the base inventory was too narrow. Those proposed tags were not accepted automatically. They went through manual review first.

Examples of approved expanded tags include bear, cat, dog, rabbit, running, pointing, thumbs_up, table_flip, flower, glasses, and gun.

In practice:

• base tags are the project's original controlled vocabulary
• expanded tags are the hand-approved additions that came out of the LLM review loop

How the data was cleaned

This section is the short version of the full pipeline.

Step 1. Load and validate the source records

Rules applied:

• the root JSON node must be an object keyed by kaomoji strings
• every record must have a non-empty kaomoji
• every record must include both original_tags and new_tags
• both tag fields must be list[str]

Step 2. Normalize and merge raw tags

Rules applied:

• Unicode normalization with NFKC
• collapse repeated whitespace to a single space
• trim surrounding whitespace
• lowercase tags
• drop empty tags
• preserve first occurrence while removing duplicates

Step 3. Remove obvious noise from tags

Rules applied:

• exact-match removal for non-semantic labels such as featured or popular
• regex-based removal for numbering and sequence-like labels such as day #4
• mark tags containing _, /, |, or multiple hyphens as normalization candidates

Step 4. Normalize tags into the base vocabulary

Rules applied:

• map variants through lemma_map
• map semantic equivalents through semantic_map
• split the result into canonical tags and unmapped tags

Step 5. Score record quality and detect duplicates

Hard filters:

• empty kaomoji
• length shorter than 2 or longer than 80
• control characters
• plain multi-token text strings

Duplicate handling:

• exact duplicate detection
• normalized duplicate detection
• near-duplicate detection with repeated-character compression

Step 6. Strict candidate selection for LLM annotation

The final LLM pass focused on rows with sparse rule tags, many unmapped tags, or structurally valid kaomoji that still looked under-described.

Step 7. LLM annotation

The annotation stage used OpenAI Batch with gpt-5.4. Each request included the kaomoji string, rule tags, normalized and unmapped tag context, the rule-based quality score, and a low-resolution rendered image of the kaomoji.

The model returned:

• primary_tags
• secondary_tags
• optional proposed_new_tag
• reject
• reject_reason
• confidence

Step 8. Manual review of proposed tags

Rejecting a proposed tag did not mean rejecting the kaomoji itself. It only meant "do not add this new label".

Step 9. Final export

Rows with reject=true are excluded from the public dataset. For the rows that remain, approved proposed tags are merged into secondary_tags, and tags is built from the union of primary_tags and secondary_tags.

Source field

The public source field is fixed to:

emoticon_dict+llm+human_review

What the scores mean

quality_score is a local rule-based score. It reflects how plausible and useful the sample looked before the LLM stage.

llm_confidence is the model's own confidence estimate for the annotation it returned.

These two numbers should not be treated as the same thing.

Known limitations

• Some tags are easier to apply consistently than others.
• Scene-like ASCII constructions are fuzzy at the boundary between "kaomoji" and "ASCII art".
• The source repository is heterogeneous. Some rows are neat single-face kaomoji; others are long action strings or memes.
• quality_score and llm_confidence are operational signals, not gold labels.

Recommended uses

Good fits:

• browsing and searching kaomoji by emotion, gesture, or motif
• building a taggable kaomoji picker
• data exploration and clustering
• retrieval systems

Less ideal fits:

• benchmark-grade evaluation without extra curation
• tasks that require perfectly stable fine-grained labels

Reproducibility and code links

The release was produced by a local pipeline in this repository. The public code links are placeholders for now and can be replaced once the code is uploaded.

• Source repository used as raw input: https://github.com/ekohrt/emoticon_kaomoji_dataset
• Project repository : https://github.com/Kare-Udon/kaomoji-labeled-tool

Citation

If you use this dataset, please cite both the original source repository and this derived dataset release.

Support

LLM labeling costs me about 50$.

If you feels this dataset is helpful, maybe considered buying me a cup of coffee.