Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?

About

The dataset contains 303 professional-work tasks with prompts, reference inputs, rubrics, expert completion times, and two sets of expert reference deliverables. Expert-only reference deliverables are the finished work produced by an expert without AI assistance; expert-with-AI reference deliverables are the finished work produced by an expert using AI. It combines 220 original GDPval tasks from OpenAI with Snorkel AI annotations and 83 GDPWorks tasks created by Snorkel AI. Scaling Intelligence publishes the dataset in collaboration with HazyResearch.

Quick start

from datasets import load_dataset

ds = load_dataset("ScalingIntelligence/joules-to-gdp", token=True)
print(len(ds["gdpval"]), len(ds["GDPWorks"]))  # 220, 83

Dataset details

Splits

Split	Tasks	Origin
gdpval	220	Original tasks by OpenAI; annotations added by Snorkel AI
GDPWorks	83	Tasks and both sets of expert reference deliverables created by Snorkel AI

The 83 GDPWorks tasks add 36 detailed O*NET occupations beyond GDPval's 44, for 80 across both splits. They have higher task difficulty: median expert-only completion time is 4.0 hours, versus 2.34 hours for GDPval.

Dataset sources

• Paper: Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?
• Code and documentation: Joules to GDP on GitHub
• Original task source: OpenAI GDPval

Files and references

Each row contains a task prompt, IDs and occupation information, expert_only_completion_time, expert_with_ai_completion_time, input reference_files, expert_only_deliverables, and expert_with_ai_deliverables. File records include a repository path or upstream source_url, original filename, size, and SHA-256; relation distinguishes declared from supplemental files. For GDPval, source_task_id maps each row to OpenAI's original task.

OpenAI input references are linked to pinned openai/gdpval revision 11e7900, not mirrored here. Download them with:

python download_gdpval_references.py --output ./openai_references

The helper requires huggingface_hub and pyarrow and verifies downloaded files against their recorded hashes. Asset paths are relative to the repository root; download a deliverable using its expert_only_deliverables[0].path with huggingface_hub.hf_hub_download(repo_id="ScalingIntelligence/joules-to-gdp", repo_type="dataset", filename=path, token=...).

Sponsors

Snorkel AI · Modal · Laude Institute · Google Cloud Platform · IBM Research · Stanford HAI · Stanford Marlowe.

License

This is a mixed-source dataset. Original Snorkel AI-created annotations, the 83 GDPWorks tasks, and their expert reference deliverables are offered under CC BY 4.0 to the extent Snorkel AI holds the necessary rights. That grant does not cover OpenAI-originated GDPval task text and original rubrics or third-party material. See the dataset license and provenance notice for the scope of the grant. The project GitHub repository's Apache-2.0 license applies to its code.

Citation

@misc{zeinali2026joulestogdpeconomicvalueai,
      title={Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?},
      author={Amirreza Zeinali and Jon Saad-Falcon and Avanika Narayan and Mersad Abbasi and Jonathan S. Hartley and John Hennessy and Christopher Re and Azalia Mirhoseini},
      year={2026},
      eprint={2610.30858},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.30858},
}