J-PragEval-v0

A minimal-pair benchmark isolating four Japanese pragmatic phenomena from surface fluency: honorific register, implicit subject reference, in-group and out-group reference, and indirect refusal.

Introduced in arxiv.org/abs/2608.14896.

What it is for

JGLUE-style benchmarks roll lexical, syntactic and pragmatic competence into a single score. This set holds surface fluency constant across each pair so the only thing that varies is the pragmatic contrast, which makes it usable for probing and for teacher-forced log-probability evaluation rather than only for scoring.

Files

File	Phenomenon	Pairs
honorifics.jsonl	honorific register	6
implicit_subject.jsonl	implicit subject reference	6
in_out_group.jsonl	in-group / out-group reference	6
indirect_refusal.jsonl	indirect refusal	6

24 pairs total. This is a v0: small enough to inspect by hand, too small for anything that needs statistical power on its own.

Fields

Field	Type	Description
id	string	Stable identifier, e.g. HON-0001
feature	string	One of the four phenomena
prompt_plus	string	Context where the marked pragmatic form is required
prompt_minus	string	Minimally different context where it is not
continuation_correct	string	Pragmatically appropriate continuation
continuation_incorrect	string	Grammatical but pragmatically wrong
notes	string	Why the contrast holds, in English
annotator_agreement	float	Currently 1.0 throughout, see below

Each item is a scenario pair rather than a sentence pair. prompt_plus and prompt_minus describe the same action in two social configurations, so the pragmatic demand changes while the propositional content does not. A model is scored on whether its preferred continuation flips with the scenario.

annotator_agreement is 1.0 on every item because the set has a single annotator. It is a placeholder for a future multi-annotator release, not a measurement.

What was measured on it

On TinySwallow-1.5B: honorific register is linearly decodable at 0.96 balanced accuracy at layer 15, with a 93% behavioural flip rate. Implicit subject and in-group reference are not linearly decodable at the final prompt token (0.48 and 0.38) despite flip rates of 0.77 and 0.79, which suggests those contrasts are resolved during generation rather than stored at the prompt.

Known limitation

The indirect-refusal pairs conflate politeness with continuation length: probe accuracy of 0.95 collapses to a 0.43 flip rate under length-normalised teacher forcing. Treat that subset as diagnostic of the benchmark rather than of the model, and see the paper before using it.

This is v0. It is small, single-annotator, and not adversarially filtered.

Citation

@article{braun2026jprageval,
  title  = {Interpretable Cross-Lingual Alignment in Small Language Models:
            Probing Cultural and Pragmatic Reasoning in Japanese-English
            Bilingual LLMs},
  author = {Braun, Florian},
  year   = {2026},
  journal= {arXiv preprint arXiv:2608.14896},
  doi    = {10.48550/arXiv.2608.14896},
  url    = {https://arxiv.org/abs/2608.14896}
}