The work

Each task starts with sourcing: you find a publicly accessible Odia PDF page in an assigned document type — a newspaper, a textbook, an exam paper, a flyer, form, manual, menu, brochure, notice or worksheet — that contains at least one multimodal element such as a table, diagram, image or handwriting, and you record where you got it. You then produce a full structural map of that page: bounding boxes around every meaningful region, a component type for each (document title, section heading, paragraph, list, table, figure, diagram, caption, formula, question, answer field), a reading-order index that reflects how the page is actually read across columns, and parent-component links tying captions and cells to the figure or table they belong to. Finally you transcribe every text region exactly as it appears in Odia script — never transliterated — flagging illegible regions rather than guessing, and you record page metadata and flags for tables, formulas and handwriting.

The dataset is deliberately weighted toward what parsing and vision-language models handle worst: dense multi-column newspaper pages, handwritten forms, mixed-script material, exam layouts with answer fields. All of it is human-authored — no parsing model output is passed through. Experienced annotators also review colleagues' tasks end to end as a second Odia expert.

What the screen looks for

The platform screen is interested in verifiable things: native Odia fluency with full command of the script, its diacritics and conjunct forms; prior hands-on document work (annotation, transcription, translation, localization, subtitling, proofreading, journalism, regional-language data review); and whether you apply a taxonomy consistently rather than improvising per document. Expect follow-ups on how you would handle specific edge cases — a caption sitting between two columns, a handwritten marginal note, a numeral in Odia versus Western form — and on whether you can distinguish a transcription defect from a stylistic choice. Nice-to-haves that get noticed: OCR correction or post-editing, MTPE, bilingual QA, Odia typesetting or digitization, and Unicode normalization and input-method fluency.

Logistics

  • Fully remote and asynchronous, open only to contributors based in India, the United States, Canada or Western Europe.
  • Observed pay on this listing is around $13/hour; bands on the platform vary by project and are not guaranteed.
  • Hours are flexible and task-based. Accuracy is the first measure on this project, with handling time tracked alongside it — so pace matters, but not at the cost of a wrong diacritic.
  • You need a computer capable of reliable PDF viewing and an Odia input method you can type in fluently.