STAIR‑Captions, released in 2017, offers 820,310 Japanese captions for caption generation, multimodal retrieval, and image generation, complete with detailed annotations, metadata, and a Creative Commons BY‑4.0 license.
Description in other languages
Statistics
Statistics as of Nov 16, 2026
Hugging Face downloads (last 30 days)
2,000
Hugging Face likes
7
Listed
Jul 1, 2024
Downloads (30-day window) over time
Show as a table
| Month | DL (30d) |
|---|---|
| Dec 2025 | 961 |
| Jan 2026 | — |
| Feb 2026 | 196 |
| Mar 2026 | 414 |
| Apr 2026 | 531 |
| May 2026 | 829 |
| Jun 2026 | 366 |
| Jul 2026 | 476 |
| Aug 2026 | 652 |
| Sep 2026 | 965 |
| Oct 2026 | 2,000 |
| Nov 2026 | 2,000 |
Downloads over the 30 days before the last snapshot of each month.
Popular in the same category
- Cauldron-JAturing-motors7kCauldron‑JA is a Japanese vision‑language dataset of 44 sub‑datasets translated from The Cauldron using the DeepL API, available via HuggingFace’s datasets library and licensed identically to the original set, with prompts released under CC‑BY‑4.0.
- Galgame-VisualNovel-Reuploadjoujiboi5kRestructured reupload of the Galgame VisualNovel dataset (OOPPEENN/56697375616C4E6F76656C5F44617461736574) for efficient Hugging Face datasets loading, preserving all original audio/text and providing an extraction script with multiple game‑subset options.
- JMMMUJMMMU3kJMMMU is a Japanese multimodal benchmark expanded over tenfold to 1,320 culturally diverse questions (720 culture‑agnostic, 600 culture‑specific) translated by native subject experts, now featuring a public leaderboard.
- ogiri-boketeYANS-official1kA Japanese‐only dataset from the Bokete crowdsourced joke site, split into three tasks—text‑to‑text, image‑to‑text, and text‑in‑image‑completion—with 500 image prompts for image‑to‑text (2355 answers) and 100 each for the other two tasks (≈900 answers each), processed via OCR and filtered for inappropriate content.
- JMMMU-ProJMMMU1kJMMMU‑Pro is an image‑based Japanese multimodal benchmark created by generating visual questions with a realistic model and human verification, revealing that current open‑source LMMs perform poorly on it and offering a cost‑effective method for future VQA benchmark development.
- vntl-leaderboardlmg-anon857A leaderboard that ranks large language models by their semantic accuracy in translating Japanese visual novels into English, using cosine similarity over 256 samples and reporting chrF scores for comparison with tools like Sugoi Translator, Google Translate, Naver Papago, and Alibaba Translate.
Resources in similar research fields
- WAONspeed281WAON is a large, high‑quality Japanese image‑text pair dataset built with size, SigLIP‑score filtering and deduplication (by URLs, captions and pHash), and released on HuggingFace under Apache 2.0 for informational analysis.Shared: Annotation and Dataset Development
hiragana_mojigazondl-lab18Text image dataset (Hiragana 73 characters version)GitHubShared: Annotation and Dataset Development
j-ono-dataObakeConstructs11A simple, open-source collection of Japanese onomatopoeic and mimetic sound words in JSON format. With manga samples.GitHubShared: Annotation and Dataset Development- Japanese-Medical-VQA-12mMIL-UT423Japanese Medical VQA 12M is a large‐scale multimodal dataset (≈12 M samples) of Japanese medical images with original and enriched captions in both source language and Japanese, plus generated VQA question‑answer pairs, released in Parquet/Webdataset format and built from Open‑PMC‑18M using InternVL3.5 for enrichment, Qwen3‑30B‑A3B for translation, and GPT‑OSS 120B for VQA generation.Shared: Annotation and Dataset Development
ja-vg-vqayahoojapan30Japanese Visual Genome VQA datasetGitHubShared: Annotation and Dataset Development- Jaglellm-jp168Jagle is a ~9.2 million‑instance Japanese multimodal post‑training dataset built from image–text pairs and PDF corpora, used to train LLM‑jp‑4‑VL 9B beta and shown to boost performance on Japanese vision‑language tasks.Shared: Annotation and Dataset Development