Specification

Corpus specification, edition 1.

Everything a technical team needs to decide: schema, closed vocabularies, taxonomy, composition, coverage, quality control, disclosures, delivery formats and licence. Print this page to keep a PDF.

1 · What it is

Trap-based evaluation and preference data.

A 3,009-item corpus for English–Arabic legal and financial translation, across 74 thematic chunks. Each item isolates one diagnosable failure mode, pairing the attorney-grade correct rendering against the plausible-but-wrong output an MT system or non-lawyer actually produces, with a written rationale explaining what the wrong version does to the legal meaning.

3,009items
74thematic chunks
15error types
4registers
5legal domains
2directions
2 · Schema

Two field sets, keyed by task.

One JSON object per line. Key order is constant. IDs run alfe-0001–alfe-3009, continuous, no gaps.

EN→AR — en_to_ar_term, clause_faithfulness, register_adaptation

{
  "id":          "alfe-0010",
  "domain":      "islamic_finance",
  "register":    "MSA",
  "task":        "en_to_ar_term",
  "source_en":   "constructive possession",
  "gold_ar":     "القبض الحُكمي",
  "chosen_ar":   "القبض الحُكمي",
  "rejected_ar": "الحيازة البنائية",
  "error_type":  "domain_term_of_art",
  "rationale":   "the established fiqh term ...",
  "difficulty": "hard"
}

AR→EN — ar_to_en_term

The same set with source_ar, gold_en, chosen_en and rejected_en in place of the four EN→AR answer fields.

Closed vocabularies

domaincontract_law · corporate · litigation · islamic_finance · M&A
registerMSA · Egyptian Arabic · Levantine Arabic · Gulf Arabic
tasken_to_ar_term · clause_faithfulness · ar_to_en_term · register_adaptation
difficultyeasy · medium · hard, skewed hard
error_typeFifteen values, listed in section 3
gold and chosen are not interchangeable. gold is the full reference, often carrying a bracketed gloss or an accepted variant; chosen is the tight form a practitioner writes. Use gold for scored evaluation and chosen for preference pairs.
3 · Taxonomy

Fifteen named failure modes.

Every item carries exactly one. Group results by it and you get a failure profile, not a score.

error_typeWhat goes wrong
polysemy_false_senseRight word, wrong semantic branch
legal_meaning_driftPlausible Arabic that loses or inverts the legal effect
near_synonym_driftAdjacent term that shifts doctrine or register
concept_calqueStructural transfer that strips the legal concept
false_cognateDiacritical or near-form confusion
domain_term_of_artWrong choice between two doctrinal bodies
literal_calqueWord-by-word where a term of art exists
under_translationRight root, too weak — strips an element
register_and_complianceRegister failure and compliance failure stacked
compliance_criticalSharia-compliance violation (riba framing and kin)
term_of_artNon-standard, academic, or wrong-tradition term
register_mismatchRight concept, wrong register for the audience
unnecessary_transliterationLatin importation where the Arabic is settled
modality_lossObligation downgraded to prediction
void_vs_voidableConflates void ab initio with rescindable
4 · Composition & coverage

What the file covers.

DimensionFull corpusPilot set (505)
Items3,009 (alfe-0001–3009)505 (alfe-0001–0505)
DomainFive domains; corporate and contract law the largestcontract_law 169 · corporate 118 · islamic_finance 89 · litigation 87 · M&A 42
RegisterMSA plus 220+ dialect items in matched triplesMSA 453 · Gulf 18 · Egyptian 17 · Levantine 17
TaskFour types; 480+ operative clauses, 240+ AR→EN itemsen_to_ar_term 295 · clause_faithfulness 81 · ar_to_en_term 77 · register_adaptation 52
DifficultySkewed hardmedium 239 · hard 171 · easy 95
Compliance-critical95+ items28 items (plus 7 register_and_compliance)

Pilot figures are machine-verified against the delivered file itself.

5 · Quality control & disclosures

What we check, and what we tell you before you buy.

Checks on every build

  • Schema and key order per task type
  • Taxonomy values within the published closed sets
  • Script integrity: no Latin in Arabic fields, no Arabic in English fields, NFC normalised
  • Contiguous IDs, no gaps
  • chosen ≠ rejected; no duplicate sources within or across builds
  • Second-pass read of every dialect and compound-clause item, plus a random sample, then author sign-off

Disclosed up front

  • Nine source headwords appear twice, all within alfe-0001–1001 and listed with IDs. Six pairs test materially different failure modes; three are near-redundant.
  • Some items describe regimes that move — virtual-asset licensing, the BNPL perimeter, apostille membership, passenger-rights compensation.
  • Dialect items are a deliberate minority: MSA carries the drafting vocabulary and is weighted accordingly.
  • Chunk notes travel with each block of 42 items, listing domain, distribution and review priorities.
6 · Delivery & licence

What you receive, and on what terms.

FormatJSONL, one object per line, plus a pretty-printed JSON review copy. Other formats — preference pairs, SFT triples, Parquet — on request at scaled tiers.
AccompanyingBuild notes, schema card and licence certificate. Per-chunk files and chunk notes at scaled tiers.
DeliveryPilot on purchase. Larger licences by private expiring link, with a 14-day acceptance window from delivery.
Permitted useTraining, fine-tuning, evaluation, benchmarking and alignment. Models and outputs belong to the Licensee.
Not permittedRedistributing, publishing or sublicensing the corpus, or using it to build a competing dataset.
Governing lawThe laws of the State of Delaware; Delaware courts.
Request the sample → See pricing Licence terms

Cite as: Fidem Praxis, “Arabic Legal & Financial Translation Corpus,” Edition 1, 3,009 items, https://fidempraxis.com/services/arabic-legal-corpus