Early accessLLMGR-LLM-04

Professional Work Tasks

Expert-graded finance, legal and office tasks with real deliverables.

Request sample Download spec (PDF)Get pricing

Domains: Finance · Law · Office
Deliverables: XLSX · DOCX · PPTX
Rubric items per task: 27–65
Languages: EN · ZH

Illustrative rendering of annotation layers.

01Overview

Open-ended professional tasks that end in real work products — spreadsheets, memos and slide decks. Each task bundles the source documents, a reference deliverable and a detailed rubric, so models can be trained and scored on the work analysts, lawyers and office professionals actually do.

Modalities

Documents
Spreadsheets
Slides
Rubrics

Use cases

Reinforcement learning on long-horizon office work
Evaluation of agents that produce documents, spreadsheets and decks
Rubric and judge-model development

02Specification

Task families

Finance: Capital, margin and ratio analyses from public filings; spreadsheet model plus written memo (English)
Law: Simulated case files based on landmark matters; memo, control matrix and board briefing (Chinese)
Office: Training and budget decks built from public source documents (Chinese)
Occupational: Multi-file tasks by sector and occupation, such as compliance and healthcare management (English)

Each task includes

Brief: Task instructions, constraints and acceptance criteria
Inputs: Source documents: PDF, DOCX, XLSX, CSV and HTML
Reference deliverable: Gold output in the requested formats
Rubric: 27–65 criteria per task, with points, numeric tolerances and partial-credit rules
Metadata: Domain, difficulty tier and targeted skills

Provenance

Inputs: Public filings and documents, or synthetic material; no personal health data
Model trajectories: Not included
Volume: Confirmed per order

03Schema

Field	Type	Description
instruction.md	markdown	Task brief given to the model.
task.toml	toml	Domain, difficulty and skill tags.
input_files/	directory	Source documents for the task.
golden_output/	directory	Reference deliverables.
rubric	json / toml	Criteria with points, tolerances and grading notes.

04File structure

Directory layout

professional-work-tasks/      # indicative; final layout confirmed per delivery
└── tasks/<domain>/<task_id>/
    ├── instruction.md
    ├── task.toml
    ├── input_files/
    ├── golden_output/
    └── tests/
        └── rubric.toml

05Sample

Sample on request.

Tasks from each domain with inputs, reference deliverables and rubrics

Size: On request
Formats: XLSX, DOCX, PPTX, PDF, JSON

Request sample Download spec (PDF)

06License

Graviti Commercial Data License

Commercial training and evaluation rights. Volume and exclusivity agreed per order.

How we license data →

07FAQ

Questions about Professional Work Tasks

What is the Professional Work Tasks dataset?

Professional Work Tasks (GR-LLM-04) is a LLM training dataset from Graviti. Open-ended professional tasks that end in real work products — spreadsheets, memos and slide decks. Each task bundles the source documents, a reference deliverable and a detailed rubric, so models can be trained and scored on the work analysts, lawyers and office professionals actually do.

What is Professional Work Tasks used for?

Typical uses: reinforcement learning on long-horizon office work, evaluation of agents that produce documents, spreadsheets and decks and rubric and judge-model development.

What modalities and annotations does Professional Work Tasks include?

Professional Work Tasks includes Documents, Spreadsheets, Slides and Rubrics. Its schema has the fields instruction.md, task.toml, input_files/, golden_output/ and rubric.