README.md

metadata

license: cc-by-4.0
language:
  - prs
  - fa
  - en
task_categories:
  - translation
pretty_name: >-
  Sawda: Intent-Preserving Translation in Trade Communication
  (Dari/Farsi–English)
size_categories:
  - n<1K

Sawda

A pilot benchmark for intent-preserving translation of culturally loaded trade communication between Afghan Dari and English. 60 native-validated fictional items; 324 translations from six open-weights systems; 180 blind human judgments by a native Dari speaker; two LLM-judge calibration sets.

Paper and code: https://github.com/MurtazaKafka/sawda-bench

Files

file	contents
data/items.jsonl	60 validated items: id, direction (fa2en/en2fa), utterance_src, context, literal_render, intent_render, pragmatic_note, failure_class, tranche (seed/lit/synth), lang_code
data/translations.jsonl	324 system outputs (54 eval items × 6 systems), verbatim, with measured token usage; 4 empty outputs flagged
data/human_judgments.jsonl	180 blind native-speaker judgments: intent (3/2/1), appropriateness (1–7), notes, quick-flags
data/llm_judgments.jsonl, data/llm_judgments_rubric_v1.jsonl	DeepSeek v4-pro judge calibration (rubric v2 and v1)
data/fable_judgments.jsonl	Claude Fable 5 panel calibration
data/heldout_ids.json	6 items held out as few-shot exemplars (excluded from evaluation)
data/judging_plan.json	The randomized blind judging order
drafts/*.jsonl	Pre-validation drafts, released for diffing against validated items

Data statement (summary)

• Language variety: the Persian side is Afghan Dari (all items carry lang_code: prs_Arab); Iranian Farsi appears only as an unwanted output register in system translations. Direction field names fa2en/en2fa are legacy labels denoting Dari.
• Provenance: 8 seed items written by a native Dari speaker from lived trade/family-business communication; 20 items grounded in the taarof and business-pragmatics literatures; 32 validated synthetic variants. All items are fictional: no real names, businesses, or identifying details.
• Annotation: a single native Dari-speaking validator/judge (the benchmark author). Judgments are blind to system identity but not item provenance. No inter-annotator agreement is available yet.
• Known limitations: pilot scale (30 judged items); machine judge labels failed calibration (κ ≤ 0.17) and are released as calibration data only, not evaluation labels.

Ethics

Dari materials involve communities under active persecution. All items are fictional. See the paper's Ethical Considerations section.

Citation

Paper draft in the GitHub repo (paper/). Formal citation forthcoming.