anna-sarvam/economic-index-structured

Economic Index - Structured & Cleaned Dataset

This dataset is a cleaned, structured version of the Anthropic Economic Index, organized for easy integration with persona-based scenario generation pipelines.

Dataset Description

The Anthropic Economic Index tracks how people use Claude AI for work-related tasks. This structured version extracts and organizes the key information into easy-to-use tables.

Original Data Period: August 4-11, 2025
Source: Anthropic Economic Index Release 2025-09-15
Processing: Extracted from enriched_claude_ai.csv with comprehensive structuring

Dataset Structure

This dataset contains 5 splits:

1. tasks (2,616 rows)

All unique tasks people do with Claude AI, with usage metrics.

Columns:

• task_name (string): Description of the task
• onet_task_count (float): Number of conversations using this task
• onet_task_pct (float): Percentage of total usage
• onet_task_pct_index (float): Specialization index
• automation_pct (float): Automation percentage (where available)
• augmentation_pct (float): Augmentation percentage (where available)
• has_automation_data (bool): Whether automation data exists
• has_augmentation_data (bool): Whether augmentation data exists
• has_usage_data (bool): Whether usage data exists

Example:

from datasets import load_dataset
ds = load_dataset("anna-sarvam/economic-index-structured")
print(ds['tasks'][0])
# {'task_name': 'write new programs or modify existing programs...', 
#  'onet_task_count': 6618.0, 'onet_task_pct': 0.52, ...}

2. collaboration_patterns (5 rows)

How users interact with Claude AI.

Patterns:

1. directive (38.8%) - Direct instructions
2. task iteration (22.2%) - Step-by-step refinement
3. learning (20.3%) - Educational assistance
4. feedback loop (10.3%) - Iterative improvement
5. validation (4.5%) - Verification

Columns:

• pattern_name (string): Name of collaboration pattern
• collaboration_count (float): Number of uses
• collaboration_pct (float): Percentage of total

3. task_collaboration_intersections (4,528 rows)

Which collaboration patterns are used for which tasks.

Columns:

• task_name (string): Task description
• collaboration_pattern (string): Pattern used
• onet_task_collaboration_count (float): Count for this combination
• onet_task_collaboration_pct (float): Percentage within task

4. occupations (22 rows)

SOC (Standard Occupational Classification) occupation groups.

Top Occupations:

1. Computer and Mathematical (35.9%)
2. Educational Instruction and Library (12.3%)
3. Arts, Design, Entertainment, Sports, and Media (8.2%)

Columns:

• soc_group (string): Occupation group name
• percentage (float): Percentage of classified tasks
• facet (string): Data facet

5. india (65 rows)

India-specific usage patterns and top tasks.

Columns:

• data_type (string): Type of data (overall_metric, top_task, collaboration_pattern)
• metric_name (string): Name of metric
• value (float): Metric value
• item_name (string): Task or pattern name (if applicable)

Key Statistics

• Total Tasks: 2,616 unique tasks
• Collaboration Patterns: 5 main types
• Occupation Groups: 22 SOC categories
• Task-Pattern Combinations: 4,528
• Geographic Coverage: 201 countries (including India)

Usage Examples

Load the entire dataset

from datasets import load_dataset

ds = load_dataset("anna-sarvam/economic-index-structured")

Get top 10 tasks

tasks = ds['tasks'].to_pandas()
top_10 = tasks.nlargest(10, 'onet_task_count')
print(top_10[['task_name', 'onet_task_count']])

Find education-related tasks

tasks = ds['tasks'].to_pandas()
education_tasks = tasks[tasks['task_name'].str.contains('education', case=False)]

Get India-specific top tasks

india = ds['india'].to_pandas()
india_top_tasks = india[india['data_type'] == 'top_task']
top_5_india = india_top_tasks.nlargest(5, 'value')

Find tasks for software developers

tasks = ds['tasks'].to_pandas()
software_tasks = tasks[tasks['task_name'].str.contains('software|program|code', case=False)]

Analyze collaboration patterns

patterns = ds['collaboration_patterns'].to_pandas()
print(patterns[['pattern_name', 'collaboration_pct']].sort_values('collaboration_pct', ascending=False))

India-Specific Insights

Usage Statistics

• Total Conversations: 1,831
• Global Percentage: 0.88%
• Automation: 45.5%
• Augmentation: 54.5%

Top 5 Tasks in India

1. Write/modify programs (6,618 uses)
2. Fix software errors (5,118 uses)
3. Adapt software to new hardware (3,594 uses)
4. Debug and correct errors (2,663 uses)
5. Build/maintain websites (2,661 uses)

Top 3 Collaboration Patterns in India

1. directive (44.7%) - Higher than global average
2. task iteration (23.4%)
3. learning (14.5%)

Use Cases

1. Persona-Scenario Matching

Match tasks from this dataset to expanded personas based on occupation:

# Load tasks
tasks = ds['tasks'].to_pandas()

# Filter for teachers
education_tasks = tasks[tasks['task_name'].str.contains('educat|teach|tutor', case=False)]

# Match to teacher personas

2. Realistic Collaboration Patterns

Use actual collaboration patterns in scenario generation:

patterns = ds['collaboration_patterns'].to_pandas()

# Sample by actual distribution
sampled_pattern = patterns.sample(1, weights='collaboration_pct')

3. India-Specific Scenarios

Generate scenarios using India's actual usage patterns:

india = ds['india'].to_pandas()
india_tasks = india[india['data_type'] == 'top_task'].nlargest(20, 'value')

Data Processing

This dataset was created from the Anthropic Economic Index through:

1. Download: Extracted enriched_claude_ai.csv (137K rows)
2. Filtering: Selected global-level data (geo_id='GLOBAL')
3. Structuring: Organized by facets (tasks, collaboration, occupations)
4. Flattening: Converted nested metrics to flat tables
5. India Extraction: Isolated India-specific patterns (3,874 rows)

Automation vs Augmentation

Global Averages:

• Automation: 51.1% (AI does the task)
• Augmentation: 48.9% (AI assists human)

India:

• Automation: 45.5%
• Augmentation: 54.5%

India shows more augmentation-focused usage compared to global patterns.

Limitations

• Data from only one week (Aug 4-11, 2025)
• Filtered for privacy (>200 conversations per country)
• "not_classified" and "none" categories removed for clarity
• Some tasks may not have automation/augmentation data

Citation

If you use this dataset, please cite both the structured version and the original:

@dataset{economic_index_structured,
  title={Economic Index - Structured & Cleaned Dataset},
  author={Your Name},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/anna-sarvam/economic-index-structured}
}

@dataset{anthropic_economic_index,
  title={Anthropic Economic Index},
  author={Anthropic},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Anthropic/EconomicIndex}
}

Related Resources

• Original Economic Index
• Economic Index Paper
• O*NET Database

License

Same as the original Anthropic Economic Index dataset (MIT License).

Maintenance

This is a snapshot of the Economic Index as of September 2025. For the most up-to-date data, refer to the original dataset.