ericssonbear/mr-right-zhtw

Mr. Right zh-TW — Traditional-Chinese Multimodal Retrieval Corpus

A bilingual (Traditional Chinese + English) multimodal document retrieval corpus: 769,245 Wikipedia-derived documents, each with an English title/body, a Traditional-Chinese (Taiwan) translation, and a machine-generated Traditional-Chinese caption of the document's image — plus test/validation query sets and the pseudo-labelled pairs used to fine-tune a retriever on it.

The documents derive from Wikipedia through Google Research's WIT collection; the 806,357-document base and its id space match the Mr. Right collection (Hsieh et al., 2022), which is itself built on WIT. Both upstreams are credited below. To our knowledge this is the first Traditional-Chinese multimodal IR corpus of this scale on the Hub.

Images are not redistributed here. Every row carries img_url (Wikimedia) and sha1, so the image tree is reproducible; see Getting the images.

Companion repositories:

• Retriever adapters: ericssonbear/qwen3-vl-emb-8b-mrright-zhtw-lora
• Pre-computed document embeddings: ericssonbear/mr-right-zhtw-embeddings

---

⚠️ Read this before you evaluate: the query_multi image trap

The upstream multi query type ships with a query image. We found that 1,992 of ~2,000 test queries use the same image file as their gold document. A completely untrained backbone scores Recall@10 = .952 on that field — it is measuring image hashing, not retrieval.

Use query_multi / query_multi_zhtw as plain text and do not feed the query image. All numbers in the companion model card are produced under this text-only protocol. If you skip this, your results will be systematically inflated and incomparable.

Second protocol note — the clean subset: 96 of the 2,047 test queries point at documents whose image URL is a dead link, so those documents are absent from the corpus and are unreachable by any system. Scoring on all 2,047 imposes a recall ceiling of .953. We report on the clean subset (gold document present in the corpus, n = 1,951).

---

Configs

config	rows	what it is
corpus	769,245	the documents (English + zh-TW text, zh-TW image caption, image URL)
queries-test	2,047	test queries, 3 types × 2 languages
queries-val	100	validation queries, same schema
queries-finetune	900	the human-annotated train split, with both Chinese translations per row
queries-test-hv	1,100	human-verified Chinese for the test queries that have one — join on id
queries-mt-en	2,147	machine-translated English query_multi for test+val (cross-lingual baseline)
pseudo-pairs-zh	32,359	pseudo-labelled training pairs (round-trip ranked)
pseudo-pairs-hard	27,055	hard-negative-mined pairs with an LLM judge score

from datasets import load_dataset

corpus  = load_dataset("ericssonbear/mr-right-zhtw", "corpus",       split="train")
queries = load_dataset("ericssonbear/mr-right-zhtw", "queries-test", split="test")

corpus schema

field	type	notes
id	int64	upstream Mr. Right document id — non-contiguous (37,112 of 806,357 dropped)
title, doc_text	string	English original
title_zhtw, doc_text_zhtw	string	Traditional-Chinese translation, Gemma 4, Taiwan-usage normalised
img_caption	string	zh-TW description of the image, machine-generated — long and structured, see below
img_url	string	Wikimedia URL of the image
caption_source	string	own (187,057) or external (582,188) — see below
sha1	string	SHA-1 of the image bytes; stable join key since id is gappy
image_path	string	<sha1[:2]>/<sha1>.<ext> — the content-addressed layout to reproduce locally

img_caption is not a one-line caption: it is a structured description, median 663 characters, with markdown section headers (**主體：**, **場景與背景：**) and bullet lists. 82% contain ** markup and 99.8% are multi-line. If you feed captions to a text encoder, budget the tokens and consider stripping the markup first.

queries-* schema

Three query types in both languages — query_img / query_text / query_multi and their _zhtw counterparts — plus id (the gold document), title and title_zhtw. id joins to corpus.id; relevance is one gold document per query.

queries-test additionally carries img_url, doc_text and doc_text_zhtw (copies of the gold document's fields); queries-val does not — join to corpus if you need them.

The English queries are human-written, collected on Amazon Mechanical Turk (workers restricted to ≥50 approved HITs and ≥80% approval rate), each required to reference both the document's text and its image, then filtered with a retrieval system plus manual review so the gold document is unambiguously the best match. Three splits, 3,047 queries: test (2,047), validation (100), finetune/train (900).

Which Chinese translation you are getting

This matters — please read it before comparing numbers.

The Chinese queries were produced by machine translation, and several independent passes exist (Google Translate, Gemma 4, and an LLM-panel round). A separate human-verification effort then corrected a subset by hand.

config	*_zhtw columns	*_zhtw_hv columns
queries-test	Gemma 4 MT	— (see queries-test-hv, covers 1,100 of 2,047)
queries-val	Gemma 4 MT	—
queries-finetune	Gemma 4 MT	human-verified, all 900 rows
queries-test-hv	—	human-verified, 1,100 rows, join on id

*_zhtw is Gemma 4 machine translation, not human-verified. It is shipped as the primary column because it is what the released retriever was trained and evaluated on — all numbers in the model card are on queries-test.query_multi_zhtw. If you want the higher-quality Chinese, use the *_zhtw_hv columns, but then you are no longer comparable to the published metrics.

Human verification changed the text on 882 of the 900 finetune queries, so the two columns are substantially different, not cosmetic variants. Retraining the deployed model on the verified pairs did not improve retrieval (see the model card's null results).

There is deliberately no "verified" file covering all 2,047 test queries: the internal _final file that looks like one falls back verbatim to raw Google Translate for the 947 unverified ids, so shipping it as uniformly human-verified would be misleading.

queries-mt-en has split (test / val), id, and query_multi_mt_en: a machine translation of query_multi_zhtw into English, produced with Qwen3-VL-8B-Instruct (greedy). It exists to measure the cost of translation quality against the human English query_multi in queries-test — spot checks found ~2.6% of these retain stray Chinese characters.

pseudo-pairs-* schema

id (gold document), query_multi_zhtw, query_multi, rt_rank_zh, rt_rank_en (round-trip rank of the gold document under the zh/en query — lower is a cleaner pair; null = not retrieved), and for hard, judge_score (LLM judge, higher is better).

---

How it was built

Wikipedia (CC BY-SA)
   └─ WIT — Wikipedia-based Image Text (Google Research, CC BY-SA 3.0)
        └─ Mr. Right collection (Hsieh et al. 2022, CC BY-SA 4.0) — 806,357 docs, English,
           first paragraph + one image per document; ids 0…806356
             └─ Traditional-Chinese translation of title + doc_text with **Gemma 4**,
                normalised to Taiwan usage (see below)
                  └─ images fetched from Wikimedia by img_url  →  SHA-1 content-addressed tree
                       └─ zh-TW description per image, generated with a Qwen3-VL model
                       └─ keep a doc iff its image resolved AND its caption passed the quality gate
                            →  769,245 docs  (this dataset)

Translation. English title / doc_text were translated to Traditional Chinese with Gemma 4, then normalised for Taiwan usage: candidate substitutions from the Ministry of Education's cross-strait vocabulary table were applied only where a Taiwan-LLaMa pass judged them contextually correct (e.g. 影集→相簿 in "family photo album" but not in "TV series"; 水平→水準 for "skill level" but not for "horizontal alignment").

Captions. Two sources, recorded per row in caption_source:

• own (187,057 docs) — generated for this project with Qwen3-VL-8B-Instruct under vLLM, greedy decoding, zh-TW instruction prompt, 256 max tokens.
• external (582,188 docs) — three caption sets produced by other members of the same lab with their own VLM pipelines, joined by document id.

Quality gate. A document is kept only if its image was successfully fetched and its caption is non-empty, contains no replacement characters, and has at least 10 CJK characters. 37,112 documents were dropped (dead image links or failed captions) — hence the gappy id.

Getting the images

The image tree is ~1.2 TB / 769,217 files, so it is not redistributed. Rebuild it from img_url and verify each file against sha1:

import hashlib, pathlib, requests
from datasets import load_dataset

UA = "your-project/1.0 (your-contact@example.com)"   # Wikimedia requires a real UA
root = pathlib.Path("images")

for row in load_dataset("ericssonbear/mr-right-zhtw", "corpus", split="train", streaming=True):
    dst = root / row["image_path"]
    if dst.exists():
        continue
    r = requests.get(row["img_url"], headers={"User-Agent": UA}, timeout=30)
    r.raise_for_status()
    assert hashlib.sha1(r.content).hexdigest() == row["sha1"], row["id"]
    dst.parent.mkdir(parents=True, exist_ok=True)
    dst.write_bytes(r.content)

Wikimedia rate-limits to roughly 1 image/s per egress IP and expects a descriptive User-Agent and honoured Retry-After. A full rebuild from one IP takes on the order of 9 days; plan accordingly. Content addressing means parallel downloaders cannot collide.

If you only need the document vectors, skip the images and use the pre-computed embeddings.

Limitations

• Captions are machine-generated and inherit VLM failure modes (hallucinated detail, OCR errors, occasional English leakage). They were filtered for form, not for factual accuracy.
• doc_text_zhtw is machine-translated at scale and not human-reviewed. All 769,245 documents were translated by Gemma 4 with an automated Taiwan-usage pass; no human read them.
• The default query translations are machine translation too. Only the 900 finetune queries and 1,100 of the 2,047 test queries have a human-verified Chinese version, and it lives in separate *_zhtw_hv columns — see Which Chinese translation you are getting.
• No natively-authored Chinese queries. Even the verified rows are corrected translations of English queries, not queries a Taiwanese speaker wrote from scratch, so they carry translationese. Results on them are not a substitute for native-speaker queries.
• The separate queries-mt-en config is raw machine translation with no human pass; spot checks found ~2.6% retaining stray Chinese characters.
• caption_source is confounded with document id range (own covers the high-id tail), so it should not be read as a controlled comparison between captioning pipelines.
• The corpus is a public encyclopedia: no personal data beyond what Wikipedia already publishes, and no human-subject research is involved.

License and attribution

Released under CC BY-SA 4.0, inherited from the upstream collections: WIT is CC BY-SA 3.0 (which permits relicensing adaptations under a later version) and Mr. Right is CC BY-SA 4.0; both derive from Wikipedia (CC BY-SA). Any redistribution or adaptation must stay under CC BY-SA 4.0 and credit this dataset and the upstreams.

Individual images remain under their own Wikimedia licenses and are not covered by this dataset's license — that is one reason only URLs are shipped here.

Upstream:

@inproceedings{srinivasan2021wit,
  title     = {{WIT}: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning},
  author    = {Srinivasan, Krishna and Raman, Karthik and Chen, Jiecao and
               Bendersky, Michael and Najork, Marc},
  booktitle = {Proceedings of the 44th International ACM SIGIR Conference on Research
               and Development in Information Retrieval},
  pages     = {2443--2449},
  year      = {2021}
}

@article{hsieh2022mrright,
  title   = {Mr. Right: Multimodal Retrieval on Representation of ImaGe witH Text},
  author  = {Hsieh, Cheng-Ping and Chen, Jui-Ting and others},
  journal = {arXiv preprint arXiv:2209.13764},
  year    = {2022}
}

Citation

@misc{mrright_zhtw_2026,
  title  = {Mr. Right zh-TW: A Traditional-Chinese Multimodal Retrieval Corpus},
  author = {Hsin-Yu Lin and Hong-Yan Huang and Shau-Yung Hsu and Song-Duo Ma and Pin-Yu Chen and Chu-Yun Chen and Wei-Te Ho and Pu-Jen Cheng},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ericssonbear/mr-right-zhtw}
}

Acknowledgements

Built by Hsin-Yu Lin, Hong-Yan Huang, Shau-Yung Hsu, Song-Duo Ma, Pin-Yu Chen, Chu-Yun Chen, Wei-Te Ho and Pu-Jen Cheng. Heavy compute ran on the NCHC 晶創 (Nano5) cluster via Slurm. Image captioning for the low-id range was contributed by lab members running their own VLM pipelines. Funded under National Science and Technology Council, Taiwan, project NSTC 115-2634-F-001-006 ("Advancing Next-Generation Frontier AI Research" programme).