Skip to content

ravyg/ate-task-ability-dataset

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

O*NET Task → Ability Mapping Dataset

A fine-grained mapping from 4,577 O*NET work tasks to the 52 O*NET human abilities each task requires, with a graded importance weight per (task, ability) pair.

This dataset supports task-level analysis of human ability requirements — enabling research on automation exposure, skill demand, and human–AI task allocation at the task level rather than the coarser occupation level.

Built as the empirical foundation for the ATES (Agentic Task Exposure Score) framework. The mapping is independently useful for labor-economics and AI-exposure research even without the ATES model.

Find this useful? Please clone it and give the repo a star — it takes one click, helps others discover the dataset, and lets us gauge interest to keep it maintained and expanded. If you use it in your work, a citation is the best thanks of all. 🙏

🚧 Status: labels released now · human validation in progress

The full task→ability mapping (4,577 tasks) is available today. Labels were produced by an LLM against the O*NET taxonomy; independent human validation is currently underway (a 200-task audit sample, reported as Cohen's κ). Results will be added to validation/ as they complete.

Want to help — and get credited? We're actively recruiting annotators. It takes ~15 min and your name goes in the Acknowledgements → reach out here. See Want your name here? below.


What's here

File Description
data/task_ability_mapping.csv Main dataset — 24,796 (task, ability) rows over 4,577 tasks
data/abilities_reference.csv The 52 O*NET abilities and their category
data/audit_sample_200.csv 200-task stratified validation sample (human re-annotation)
validation/ Human annotations + inter-rater agreement (Cohen's κ)
crowdsource/ The web form used to collect human validation labels
docs/METHODOLOGY.md How the mapping was built and validated
docs/SCHEMA.md Column-by-column schema

Main dataset schema

data/task_ability_mapping.csv — one row per (task, ability) requirement:

Column Type Description
task_id int O*NET task identifier
occupation string O*NET occupation title the task belongs to
task_text string Full task statement
ability_name string One of the 52 O*NET abilities
weight int Importance: 1 = mild, 2 = moderate, 3 = critical
uncertain 0/1 1 if the annotation was flagged low-confidence

Coverage: 4,577 tasks · 24,796 mappings · avg 5.4 abilities/task. Weight distribution: 1 → 12.9% · 2 → 63.5% · 3 → 23.7%. Uncertain: 869 rows (3.5%) across 687 tasks.


How it was built (short version)

  1. LLM pass — each task labeled with its required abilities + weights against the 52-ability O*NET taxonomy, cross-checked against the O*NET GWA (Generalized Work Activities) chain baseline.
  2. Human validation (in progress) — a 200-task stratified sample (all uncertain-flagged tasks + random draw, seed 42) is being independently re-annotated by domain experts and volunteers.
  3. Agreement — inter-rater and human-vs-model agreement, reported as Cohen's κ in validation/. A pilot of 100 tasks reached κ = 0.81 ("almost perfect"); the full 200-task audit is underway and results will be posted as they complete.

Full protocol: docs/METHODOLOGY.md.


Source & license

  • Source taxonomy: O*NET 30.2 (U.S. Department of Labor, public domain).
  • This dataset: licensed CC-BY-4.0 — free to use with attribution.

Citation

If you use this dataset, please cite the accompanying paper. Copy-paste in your preferred style:

APA

Gupta, R., & Kumar, S. (2026). Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption. arXiv preprint arXiv:2604.00186.

MLA

Gupta, Ravish, and Saket Kumar. "Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption." arXiv preprint arXiv:2604.00186 (2026).

Chicago

Gupta, Ravish, and Saket Kumar. "Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption." arXiv preprint arXiv:2604.00186 (2026).

Harvard

Gupta, R. and Kumar, S., 2026. Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption. arXiv preprint arXiv:2604.00186.

Vancouver

Gupta R, Kumar S. Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption. arXiv preprint arXiv:2604.00186. 2026 Mar 31.

BibTeX

@article{gupta2026agentic,
  title   = {Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption},
  author  = {Gupta, Ravish and Kumar, Saket},
  journal = {arXiv preprint arXiv:2604.00186},
  year    = {2026}
}

To cite the dataset specifically, see CITATION.cff.

Acknowledgements

Special thanks to our annotators (TBD — names to be added on publication), whose independent expert re-annotation of the validation sample made the quality assessment (Cohen's κ) of this dataset possible. Their careful judgments on task→ability requirements are the backbone of its reliability.

🙋 Want your name here? Become an annotator

This dataset gets better with more human eyes — and we credit every contributor. If you'd like to help validate task→ability labels (it takes ~15 minutes, no special background needed) your name will be listed in the Acknowledgements of this dataset and its accompanying paper.

It's a simple, citable way to contribute to open research on AI and the future of work — something concrete to point to on your CV, LinkedIn, or Google Scholar.

👉 Interested? Reach out: ravishgupta.me/#contact

Tell us roughly how many tasks you'd like to annotate and we'll send you a link. Every verified contributor is acknowledged by name (or kept anonymous on request).

Related

  • Paper: Gupta & Kumar (2026), Agentic AI and Occupational Displacement, arXiv:2604.00186.
  • ATES framework — the model this dataset was built for (link on paper release).

If this dataset helped your work, please star the repo and cite the paper. It's the simplest way to support open research and keep this dataset growing. Thank you!

About

Fine-grained mapping of 4,577 O*NET work tasks to the 52 human abilities they require, with graded weights. Foundation for the ATES framework. CC-BY-4.0.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages