uratori-ja-eval

uratori-ja-eval is a Japanese evaluation set for decision models: models that take a text state and a typed question and return a probability distribution over the allowed answers. It covers three jobs: checking whether a claim or an answer is supported by supplied evidence (grounding), comparing two documents (contradictions, meaning-changing edits, added claims) and judging retrieval results and answers in a RAG pipeline (relevance, sufficiency, faithfulness). It has 2,002 items in four splits, and a further 398 items are held out privately.

Each item is a state (one or more named Japanese texts), a question, a list of options and an answer. Question types are noul (yes/no), choice (one of 2 to 7 labelled options) and score (an ordinal scale). The format mirrors the questions object of TypeSafe's /v1/systemone API, so any model or service that speaks that format can be evaluated directly. A Japanese summary is at the end of this card (日本語の説明は末尾にあります).

Splits

Split	Items	Purpose	noul / choice / score	Flagged (judges disagreed)
test	802	Reporting. A further 398 items from different document families are held out and not released.	319 / 337 / 146	96
challenge	300	Hard items only: instructions to the grader hidden in the text, unusual registers (chat, memos), question templates unseen in the uratori training data, longer documents	111 / 137 / 52	25
calibration	600	Fitting temperatures and thresholds	244 / 225 / 131	73
dev	300	Development	115 / 119 / 66	33

Tasks per split (domain field): test has 346 grounding, 175 comparison, 198 RAG, 71 writing-requirement and 12 routing items; the other splits have similar proportions. Documents do not overlap between splits; items made from the same document share a family_id. 472 items form minimal pairs: two items that differ in one detail and have different answers (pair_of).

Inputs are short to medium: the median state is about 220 tokens and the longest about 750 (ModernBERT-Ja tokenizer).

Usage

from datasets import load_dataset

ds = load_dataset("tokimoa/uratori-ja-eval", split="test")
row = ds[0]
state = {s["name"]: s["text"] for s in row["state"]}
labels = [o["label"] for o in row["options"]]

To build a /v1/systemone-style question from a row: instructions is row["question"]; for choice, criteria is {label: description} over options; for score, criteria is the list of option descriptions in order; for noul, criteria is {"true": ..., "false": ...} taken from the two options. An evaluation script that does this and computes accuracy, macro F1, ECE and family-bootstrap confidence intervals is in the uratori repository (scripts/eval_hub.py).

Fields

Field	Type	Description
id, family_id	string	Item id and the id of the document family it was made from
type	string	noul, choice or score
state	list of {name, text}	The input texts. name is referenced from the question as `name`
question	string	The question (the instructions of the API)
options	list of {label, description}	Allowed answers. noul options are false, true in that order; score options are levels from lowest to highest
answer, answer_index	string, int	The label of the correct option and its index in options
agreement	string	unanimous if the four labellers agreed, split otherwise
domain	string	grounding, compare, rag, writing or routing
difficulty	string	clear, hard or ambiguous, as intended by the author of the item
phenomena	list of string	Difficulty-inducing features present in the item: numeric paraphrase, missing information, exceptions, scope-limiting words, relative dates, decided-vs-considered, prompt injection, and others
template	string or null	Id of the question template, or null for free-form questions
pair_of	string or null	Id of the minimal-pair partner, if any

How the data was made

Documents and questions. All documents are fictional business texts (contracts, internal rules, FAQs, support tickets, meeting notes, product descriptions) about invented organisations and people, drafted with an LLM (Claude) from written guidelines that list the phenomena to include. No real documents and no text from existing datasets are included. Where an item has a free-form question (template is null), the question was written for that item rather than taken from a template.

Labels. Labels are not human annotations. For test, challenge and calibration, each item was judged independently by three LLMs (DeepSeek V4.1 Flash, Gemini 3.8 Flash, GPT-6 Luna) and the label is the majority vote among those three judgments and the answer intended by the draft; ties go to the intended answer. Items on which the four did not agree are kept and marked agreement: split, so users can report results with and without them. For dev, the label is a single LLM's judgment reviewed by the author, with the same flag for disagreement among a wider set of judges. A fourth judge (Claude Haiku 4.5) was run for reference only and does not influence the labels.

Validation. Every item passes a schema check (option counts, label membership, no duplicated content across items or splits). Phone numbers, e-mail addresses and organisation names were checked to be fictional.

Considerations for using the data

• Labels are LLM majority votes. On the flagged items the label itself is uncertain; a strong model's accuracy on them is around 0.65, against 0.94 on the unanimous items.
• The LLMs that produced the labels will score higher on this set than an independent judge would; their own accuracy (0.90 to 0.97) should not be read as a ceiling for other systems.
• The documents are fictional and were written to be tricky. Scores here will not transfer directly to any particular real document collection.
• Inputs are at most about 750 tokens. The set does not measure behaviour on long documents.
• The set is Japanese only and mostly business text.

Baselines

Accuracy on test (802) and challenge (300). Models other than uratori were run through the inference path published by their authors (compatible server, package or reference script) on 2026-10-05, with the state passed as a dict and criteria passed natively. Several of them do not target Japanese or grounding; they are listed to show what the set measures, not to rank them.

Model	Parameters	test	challenge
Jev 1.13.0 (TypeSafe API)	undisclosed	0.903	0.900
tokimoa/uratori-ja-9b v1.1	8.0B	0.897	0.913
tokimoa/uratori-ja-4b v1.1	4.2B	0.872	0.870
Mapika/decider-4b v2.1	4.2B	0.769	0.763
tokimoa/uratori-ja-2b v1.1	1.9B	0.781	0.797
Cloudflare/clef-flash	9.4B	0.747	0.800
jaredpalmer/kev-4b	4.7B	0.732	0.740
tokimoa/uratori-ja-310m v0.2	0.31B	0.686	0.693
open-alternative-jev on Qwen3.5-4B, no training	4.7B	0.621	0.583
Most frequent label per question type and option count		0.446	0.430
convaiinnovations/laya-multilingual	0.32B	0.415	0.397
Random		0.387	0.379

If you evaluate another model on this set, we would like to hear about it: open an issue at github.com/tokimoa/uratori.

License

CC BY 4.0. Please credit "tokimoa, uratori-ja-eval".

Citation

@misc{uratori-ja-eval2026,
  title  = {uratori-ja-eval: a Japanese evaluation set for grounding checks, document comparison and RAG judgments},
  author = {tokimoa},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tokimoa/uratori-ja-eval}
}

日本語

文章と質問を渡して、真偽（noul）、選択（choice）、段階評価（score）で答えさせる日本語の評価セットです。与えた根拠に主張が支持されるかの検証、2 つの文書の比較、RAG の判断（検索結果の関連性と十分性、回答の忠実性）を対象にしています。公開分は 2,002 問（test 802、challenge 300、calibration 600、dev 300）で、このほかに非公開の 398 問があります。質問の形は TypeSafe の /v1/systemone API の questions と同じにしてあるので、同じ形式を受け付けるモデルやサービスをそのまま測れます。

文書は架空の組織と人物を題材に LLM（Claude）で下書きしたもので、実在の文書や既存のデータセットの文章は含みません。正解は人が付けたものではなく、test、challenge、calibration は 3 つの LLM（DeepSeek V4.1 Flash、Gemini 3.8 Flash、GPT-6 Luna）の判定と下書きの想定の 4 者の多数決で決めました。4 者が一致しなかった件は除かずに agreement: split の印を付けて残しています（test で 96 問）。dev は 1 つの LLM の判定を作成者が確認したものを正解とし、ほかの判定者と割れた件に同じ印を付けています。

使うときの注意。割れた件では正解そのものが揺れています（強いモデルでも 0.65 前後、一致した件では 0.94）。正解を決めた LLM 自身の正答率（0.90〜0.97）は、ほかのシステムの上限として読まないでください。文書は難しくなるように書いた架空のもので、特定の実在の文書群での精度をそのまま表すものではありません。入力は最長で約 750 トークンで、長い文書での挙動は測れません。

評価のスクリプトは github.com/tokimoa/uratori の scripts/eval_hub.py にあります。ライセンスは CC BY 4.0 です。