Dataset Card for dic-nicovideo

dic.nicovideo is a comprehensive crawl and processed dump of the NicoNico Dictionary, a Japanese user-generated encyclopedia encompassing internet culture, memes, anime, and detailed event logs. This dataset provides a unique glimpse into contemporary Japanese digital subculture.

Dataset Details

Dataset Description

This dataset contains articles scraped from the NicoNico Dictionary. It includes a wide range of content, from formal definitions to hyper-detailed, community-categorized explanations of specific events, memes, and personalities prominent in Japanese internet history. The primary goal of this dataset is to provide a rich source of Japanese-language and culturally relevant text for large language model (LLM) pre-training and fine-tuning. The raw crawl data is available alongside a lightly filtered version designed for immediate use in model training pipelines.

• Curated by: KaraKaraWitch
• Funded by: featherless.ai
• Shared by: KaraKaraWitch
• Language(s) (NLP): Primarily Japanese (ja)
• License: Scripts folder is licensed under Apache 2.0. The data itself is provided under the same terms as the source (dic.nicovideo.jp). Please refer to the Licensing Information for more details.

Dataset Sources

• Repository: huggingface.co/datasets/KaraKaraWitch/dic-nicovideo

Uses

Intended Uses

As you probably guessed, the primary use case for this dataset is Large Language Model (LLM) Training. More specifically, this dataset is excellent for:

• Pre-training: Mixing this content into a pre-training corpus alongside sources like Wikipedia provides a unique, encyclopedic view of modern Japanese and internet subculture.
• Cultural Fine-tuning: Using this dataset to fine-tune a model to understand and generate text related to Japanese internet slang, memes, and cultural context.

The markdown folder contains unfiltered data for researchers who want to apply custom filtering. The jsonl directory contains the processed version of the dataset, ready for direct use in training frameworks like Axolotl.

Out-of-Scope Use

The dataset is not designed for, and performs poorly in, the following scenarios:

• Training Multimodal Models: While the source contains images, this dataset is text-only. For text-and-image applications, one should crawl NicoNico directly.
• High-Stakes Factual Q&A: This is a dictionary of internet culture, not a formal encyclopedia. It contains bias, rumors, and community-judged "facts" and should not be used as a sole source of truth for critical applications.

Misuse and Malicious Use: Frankly, it's hard to weaponize a dictionary of memes, but one could try. The dataset could potentially be used to identify and harass individuals mentioned in articles. If you do that, you're a dick. It could also be used to generate degenerate content, but let's be honest, that ship has sailed long ago with other, larger datasets.

Dataset Structure

The dataset is provided in JSONL (newline-delimited JSON) and is ready for ingestion by standard data loading pipelines (the main text field is text).

Here is a quick sample of a data instance:

{
    "title": {
        "display": "外外外外外外外",
        "category": "単語",
        "yomi": "ガイカクニナゲツヅケタラギャクテンサレタ"
    },
    "text": "外外外外外外外\n(ガイカクニナゲツヅケタラギャクテンサレタ)\n----------------------------------------\n\n## 概要\n\n2021年3月26日のプロ野球開幕戦、中日ドラゴンズ対広島東洋カープ1回戦で発生。\n\n8回表、4点の援護を受けた広島先発の大瀬良大地は一死満塁からタイムリーと内野ゴロで3点を失い、尚も二死三塁のピンチを作り降板。代わったケムナが中日の4番、ダヤン・ビシエドと対峙する。\n\n會澤は長打を警戒してか外角高めを執拗に要求し、ケムナもそれに応えてストライクゾーンを外した球を6球中5球(残り1球も外角高め)投げる。一方のビシエドもファウルで粘り、コーディエ対雄平を彷彿とさせる勝負が繰り広げられた。\n\nそして7球目、ビシエドは**外角高めに大きく外れた**150km/hストレートを流し打ち。打球は弾丸ライナーでライトスタンドへと消えていった。流石に7球も同じコースに投げたら対応されることを身を以て証明したのである。\n\nこれでリズムを崩したのか続く9回も島内颯太郎が中日打線に捕まり2失点。裏に2点返すも及ばず、中日キラーの大瀬良で開幕戦を落としてしまった。",
    "meta": {
        "rev_id": 2901188,
        "article_id": 5617458,
        "url": "https://dic.nicovideo.jp/a/%E5%A4%96%E5%A4%96%E5%A4%96%E5%A4%96%E5%A4%96%E5%A4%96%E5%A4%96",
        "length": {
            "all": 397,
            "lists_tables": 0
        },
        "sections": [
            "概要",
            "関連動画",
            "関連項目",
            "掲示板"
        ],
        "link_data": {
            "auto": true,
            "links": {}
        },
        "time": {
            "created": "2021-03-27T13:50:14+09:00",
            "updated": "2021-03-27T19:13:05+09:00"
        },
        "appraisal": {
            "clap": 0
        }
    }
}

Notably the text is formatted like so:

<TITLE>
<YOMI / Reading (if present)>
-----------------------------

<Page contents go here>

Data Collection and Processing

Data Collection: The dataset was sourced from a full web crawl of dic.nicovideo.jp. The raw, unprocessed crawl data is available in the markdown directory of this repository release for researchers who wish to apply their own filtering logic. If the HTML is needed, it is available at the folder aptly named html.

Data Processing and Filtering: The primary data distributed in this repository (jsonl) has undergone a series of light filtering stages during processing to improve data quality. The processing script handles malformed or missing entries and applies the following exclusion filters sequentially:

1. Short-form Article Pruning: Articles with very low character counts (<250 characters) are removed. These are typically stub articles, disambiguation pages, or broken entries that offer minimal informational value.
2. Listicle-based Filter: A heuristic was created to remove articles that are overwhelmingly composed of lists and tables.
   • Rationale: An excessive ratio indicates low-value or machine-generated content, not an in-depth explanatory article.
   • Mechanism: This is a super unscientific filter (as stated in the code). A threshold of 76.30% was derived by plotting the listicle_ratio and identifying the quantile (approximately the 75th) where the ratio begins to increase sharply. Articles exceeding this value are excluded.
3. Long-form "Zero-Engagement" Filter: To filter out lengthy but potentially unhelpful entries (e.g., spam, boilerplate), any article exceeding REJECT_LENGTH that has received zero positive engagement (zero "claps") is removed.
   • Rationale: A lengthy, quality article should have garnered at least some user validation over time. A complete lack of such signals suggests the entry is not a valuable contribution.
4. Niche Video Content Filter: A specific rule to target low-quality video-related entries. Articles for NicoNico videos (dic.nicovideo.jp/v/) that are both short (<REJECT_LENGTH) and have zero claps are removed, as they often represent unpopular or transient entries.

After filtering, final minor cleanup steps, such as stripping extensive link data, are performed. The resulting dataset is provided as a clean JSONL file.

Who are the source data producers?

The data is produced by the anonymous user community of the NicoNico Dictionary.

Considerations for Using the Data

Social Impact of Dataset

This dataset captures a specific and valuable niche of Japanese digital culture. LLMs trained on this data will have a better grasp of internet memes, slang, and community-specific events. However, they may also inherit the biases, inside jokes, and non-factual narratives present in a user-generated encyclopedia. It's a cultural record, not an academic one.

Discussion of Biases

The dataset contains all the biases of its source: a user-driven internet community. This includes:

• Cultural Bias: A heavy focus on anime, video games, and Japanese internet-centric topics.
• Demographic Bias: The perspectives are likely skewed towards the demographic that frequents NicoNico.
• Factual Bias: Information is crowd-sourced and may contain rumors, speculation, or intentionally misleading "joke" articles presented as fact. Explicit fact-checking was not performed.

Other Known Limitations

• Heuristic Filtering: The filters are based on unscientific heuristics and may have incorrectly removed valuable articles or retained low-quality ones. User discretion is advised.
• Scope: As a snapshot of one specific community dictionary, it is not a representative sample of the Japanese language or culture at large.

Licensing Information

The collection and transformation scripts are licensed under Apache 2.0.

Regarding the data: We do not claim any copyright over the data itself; that belongs to the individual contributors on dic.nicovideo.jp. The data is provided "as-is." Users are responsible for ensuring their usage complies with the terms of service of the original website and any applicable copyright laws. Use this dataset at your own legal risk.

Citation

If you use this dataset in your research, please cite it as follows:

@dataset{karakarawitch_dic_nicovideo,
  author    = {KaraKaraWitch},
  title     = {dic-nicovideo: A Crawl and Dump of the NicoNico Dictionary},
  year      = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/KaraKaraWitch/dic-nicovideo}}
}