OrenSchmerling/pragmatic-mt-eval

Pragmatic MT Benchmark (Hebrew, Gujarati, Vietnamese)

Dataset Summary

This is a reference-based, pragmatics-focused evaluation benchmark for Hebrew (he), Gujarati (gu), and Vietnamese (vi), with English source sentences. The benchmark evaluates pragmatic meaning in machine translation across seven dimensions: formality, speaker gender, listener gender, dialect, pronoun tier, kinship pair, and relationship.

Each base item appears in multiple language-specific variants that differ along one or more pragmatic axes. Bucket C items include conflict scenarios where the cued source text conflicts with an override condition, testing the model's ability to resolve pragmatic tensions.

This is not a general MT benchmark. It is specifically designed to probe pragmatic fidelity—whether translations preserve pragmatic nuance rather than just propositional content.

Languages

Language	Language Code	Variants
Hebrew	he	192
Gujarati	gu	288
Vietnamese	vi	240
Total		720

All source sentences are in English. Each language has variants in tracks family_and_relatives and culture_and_daily_life, distributed across buckets A, B, and C.

Dataset Structure

Files

File	Description
variants.jsonl	Flattened, viewer-friendly export: one JSON object per variant (720 lines). Contains all pragmatic fields at top level (formality, speaker_gender, listener_gender, dialect, pronoun_tier, kinship_pair, relationship, conflict_override, anchors fields, gold_reference).
eval_dataset.json	Canonical compiled dataset: a single JSON object {"variants": [...]} with nested coarse/granular/anchors/conflict_override objects. Byte-identical to data/compiled/eval_dataset.json.
he_curation.csv	Hebrew curation sheet (192 rows)
gu_curation.csv	Gujarati curation sheet (288 rows)
vi_curation.csv	Vietnamese curation sheet (240 rows)
README.md	This dataset card

Variant Fields

Field	Type	Description
variant_id	string	Unique identifier (e.g., cul_a_01-gu-level_1_intimate-female)
item_id	string	Base item identifier
source_en	string	English source sentence
lang	string	Target language code (he, gu, vi)
track	string	Topic track (family_and_relatives, culture_and_daily_life)
bucket	string	Difficulty bucket (A, B, C)
is_cued_source	boolean	Whether the source includes a pragmatic cue
formality	string	Formality level (default, formal, level_1_intimate, level_2_polite, level_3_honorific)
speaker_gender	string	Source speaker gender (male, female)
listener_gender	string	Target listener gender (male, female, null)
dialect	string	Regional dialect (northern, southern, null)
pronoun_tier	string	Second-person pronoun/register (tu, tame, aap, anh_em, chi_em, con_ong, chu_chau, co_chau, null)
kinship_pair	string	Kinship pair (null, or specific pair like mother-father, etc.)
relationship	string	Social relationship (null, or description like parent-child, etc.)
conflict_override	object|null	Conflict condition for bucket C cued items; null otherwise
must_contain_any	array[string]	Terms that must appear in any valid translation
must_not_contain	array[string]	Terms that must not appear
synset_equivalents	object	Synset-equivalent term mappings
acceptable_synonyms	array[string]	Additional acceptable synonym terms
gold_reference	string	Gold reference translation

Nested canonical schema (eval_dataset.json)

The eval_dataset.json file is a byte-for-byte copy of data/compiled/eval_dataset.json. Its structure is a single JSON object {"variants": [...]} where each variant carries nested objects under coarse (containing formality, speaker_gender, listener_gender, dialect), granular (containing pronoun_tier, kinship_pair, relationship), conflict_override (nested object or null), and anchors (containing must_contain_any, must_not_contain, synset_equivalents, acceptable_synonyms). The variants.jsonl file, by contrast, flattens all of these top-level fields for viewer-friendly per-variant JSON objects.

Evaluation Tiers

Tier	Name	Description
1	Tier1DefaultZeroShot	Free translation with no pragmatic instruction. Reveals default register, gender bias, and dialect bias.
2A	Tier2ACoarseParams	Explicit formality, speaker gender, listener gender, and dialect parameters.
2B	Tier2BGranularParams	Fine-grained pronoun tier, kinship pair, and relationship constraints.
2C	Tier2CAdversarialConflict	Source cues conflict with API parameters; the parameters must override the cues.
3A	Tier3AFewShotK5Pos	Five positive in-language examples in the prompt.
3B	Tier3BFewShotK20Pos	Twenty positive examples for stronger in-context adaptation.
3C	Tier3CFewShotK5Contrastive	Five contrastive example pairs (correct target plus an explicitly rejected alternative).

Usage

Loading with HuggingFace datasets

from datasets import load_dataset

# Load variants
ds = load_dataset("OrenSchmerling/pragmatic-mt-eval", name="variants", split="train")

# Load curation sheets
he_ds = load_dataset("OrenSchmerling/pragmatic-mt-eval", name="curation", split="he")
gu_ds = load_dataset("OrenSchmerling/pragmatic-mt-eval", name="curation", split="gu")
vi_ds = load_dataset("OrenSchmerling/pragmatic-mt-eval", name="curation", split="vi")

Evaluation Mapping

To use this benchmark with an evaluation harness:

1. Load the dataset with load_dataset(...).
2. For each variant, generate a translation from source_en into the target language.
3. Check the translation against anchors.must_contain_any, anchors.must_not_contain, etc.
4. Report per-tier and per-language accuracy.

Provenance & Construction

• Base items: 36 base source sentences
• Script: provenance/generate_sheets_template.py generates the curation templates
• Compiler: Compiles templates into data/compiled/eval_dataset.json
• Preparation: provenance/prepare_hf_dataset.py exports this dataset card and JSONL/parquet artifacts

Intended Use & Limitations

Intended Use

• Evaluating pragmatic meaning preservation in MT for he/gu/vi
• Testing formality, gender, dialect, and pronoun register handling
• Conflict resolution in bucket C cued items

Limitations

• Only three target languages (Hebrew, Gujarati, Vietnamese)
• Source language is English only
• 36 base items (720 variants total)—small-scale benchmark
• Pragmatic tiers are limited to those annotated in the data

Licensing

MIT License

Citation

@misc{pragmatic-mt-benchmark,
  title = {Pragmatic MT Benchmark for Hebrew, Gujarati, and Vietnamese},
  author = {Pragmatic MT Eval contributors},
  year = {2026},
  url = {https://github.com/pragmatic-mt-eval/pragmatic-mt-eval},
  note = {Reference-based, pragmatics-focused evaluation benchmark}
}