Home

Data

Research files, documentation, and validation sources.

Files

Replication package

i
UnitPaperGitHub
Contents

Code, retained labels, paper sources, prompts, data dictionaries, and rationale-analysis materials used to reproduce the paper. The observed-use comparison draws on public OpenAI Signals data.

GitHub

Data dictionary

i
UnitFields25 KB
Contents

Field names, descriptions, files, units, ranges, and enum values.

JSON

Task-country exposure labels

i
UnitTask x country2.96M rows
Contents

Public hierarchy-path panel. The paper reports 2.33M unique task-country labels; this panel has repeated task-country rows where tasks appear under multiple O*NET hierarchy paths.

Parquet shardsi

Split into four compact Parquet files. A CSV.GZ version would need more shards under the website asset limit.

More files

Aggregated country, occupation, industry, skill, network, and website index files.

Country summary

i
UnitCountry124 rows
Contents

Exposure, channels, margins, AI fields, and covariates.

CSV

Country-occupation table

i
UnitCountry x occupation114,452 rows
Contents

Exposure and margin shares by O*NET occupation group.

CSV.GZ

Country-industry table

i
UnitCountry x industry51,832 rows
Contents

Exposure and margin shares by retained ISIC activity.

CSV.GZ

Country-skill table

i
UnitCountry x skill778,100 rows
Contents

Exposure and margin shares by ESCO skill.

CSV.GZ shards

Split into two compressed CSV files to keep each download within the website asset limit.

Task-skill network

i
UnitNode / edge28,388 nodes / 56,878 links
Contents

CSV bundle for task-skill nodes and retained links.

ZIP

Website support indexes

Task index

i
UnitTask23,851 rows
Contents

Search fields, exposure ranks, and channel summaries.

JSONCSV

Skill index

i
UnitSkill4,537 rows
Contents

Skill families and network exposure fields.

JSONCSV

Use guide

How to use these files

Task-level exposure without a country

UseTask-country exposure labels

Aggregate the task-country rows by task. Equal-country weights describe the typical country; population weights describe people globally; GDP weights describe the world economy. A context-free benchmark answers a different question because it removes country context.

Country rankings, maps, or regressions

UseCountry summary

Use the country file for exposure, channel, margin, AI fields, and country covariates. This is the lightest file for country-level work.

Occupation or industry comparisons

UseCountry-occupation table or country-industry table

Use these for country by work-group comparisons. They are already aggregated from the task-country labels into occupation and ISIC activity cells.

Custom aggregation or validation

UseTask-country exposure labels plus the data dictionary

Start from the task-country file when the unit, weights, or sample restrictions differ from the website defaults. The dictionary gives the field meanings and allowed values.

Skills and reskilling-style analysis

UseCountry-skill table and task-skill network

Use these for skill exposure and task-skill links. This layer extends the Atlas beyond the core paper measures.

Preview

Loading preview

A few rows from the research file.

Loading preview.