This dataset contains the development tasks and grading verifiers used to evaluate the Physical-Engineering’s Last Exam (PLE) benchmark. Access is gated to protect benchmark integrity.

By requesting access you agree to ALL of the following. Misuse may result in access revocation and exclusion from the benchmark leaderboard.

Physical-Engineering’s Last Exam (Dev Set)

Physical-Engineering’s Last Exam (PLE) is a benchmark of long-horizon engineering tasks built by licensed practicing engineers across 4 disciplines. In each task, an agent works inside a realistic project workspace containing drawings, specifications, spreadsheets, solver scripts, and CAD/BIM or simulation tools to produce engineering deliverables: code, calculations, models, schedules, and technical reports.

This repository contains the 29-task public development set (split: dev) of the 85-task PLE benchmark. Each task is packaged as a self-contained Harbor task folder under tasks/<task_id>/. Every task runs with a uniform 6-hour agent time limit and no internet access beyond model API calls, and is graded on a 0 to 1 scale in an isolated verifier container (12 tasks checks; 14 tasks rubric; 3 tasks checks-judge).

Paper: arXiv:2610.30537
Benchmark site: https://ple-bench.com
Competition kit, with the evaluation runner and sample submissions: github.com/google-research/ple

Benchmark disciplines

Discipline	Benchmark tasks	Dev tasks	What the tasks cover
Structural engineering	10	4	Calculate gravity, wind, and seismic loads, verify structural members and connections, and design roof decks, diaphragms, and equipment supports.
Civil engineering	18	9	Lay out site grading, drainage, and utilities, evaluate soil settlement and foundation strategies, and verify excavation shoring and construction plans.
Electrical engineering	27	9	Analyze power distribution from utility service to branch circuits, size feeders and cables, review solar interconnections, and check drawings for energization.
Mechanical engineering	30	7	Design and troubleshoot HVAC, thermal, and fluid-flow systems, route clash-free building services, and validate equipment performance and fire simulations.

Leaderboard

A model's score is its mean reward over the 29 dev tasks; the results we report average 3 runs. The dev-set leaderboard covers every model we evaluated. Hub leaderboards for this dataset list open-weight models only, because the Hub reads results from model repositories.

Running the dev set

With the PLE evaluation runner (recommended)

To evaluate an agent harness or prepare a PLE Harness Competition entry, use the runner in the competition kit: uv run python eval/run.py <submission>. It downloads this dataset's latest commit (recorded in results.json), applies the 6-hour agent time limit, builds task-specific verifier images, grades each task in its verifier container, and exports artifacts/ for submission. While the dataset is gated, request access on this page and set HF_TOKEN to a token of the account that was granted access.

With standalone Harbor

12 of the 29 tasks also run with Harbor alone (uv tool install harbor==0.20.0) and Docker. Tasks with verifier_image = "task" under [metadata] in task.toml are graded by a verifier built from the task's own image plus tests/, which Harbor does not build, so run them with the kit's eval/run.py. In this release, that applies to the other 17 tasks: data-hall-mep-routing, fire-test-validation-memo, data-hall-mep-routing-offset, landscape-sheet-revision, data-center-site-model, engineers-estimate-check, footing-settlement-verification, solar-supplemental-review, solar-witness-test-package, fire-solve-convergence, les-mesh-resolution, fill-compaction-acceptance, shoring-design-verification, emergency-spillway-check, pre-application-capacity-response, permission-to-operate-memo, backfeed-energization-reply.

harbor run --repo fetches tasks with git clone and git lfs, so install Git LFS, and while the dataset is gated give git your Hugging Face token, for example with git config --global credential.helper store and then hf auth login --add-to-git-credential.

Of those 12, the programmatic tasks (checks) run directly from the Hub:

export ANTHROPIC_API_KEY=<your key>
harbor run --repo https://huggingface.co/datasets/google/physical-engineerings-last-exam -i frame-extension-method \
  -a claude-code -m anthropic/claude-opus-5-5 --allow-agent-host api.anthropic.com

The judge-graded ones (checks-judge) require downloading the dataset locally so you can allow your judge endpoint in task.toml (hf download uses HF_TOKEN or your hf auth login token):

hf download google/physical-engineerings-last-exam --repo-type dataset --local-dir ple
# In ple/tasks/eor-review-corrections/task.toml, under both [verifier] and [verifier.environment]:
#   network_mode = "allowlist"
#   allowed_hosts = ["<your judge's host>"]
export JUDGE_BASE_URL=https://<your judge's host>/v1 JUDGE_API_KEY=<your key> JUDGE_MODEL=<model name>
harbor run -p ple/tasks/eor-review-corrections -a claude-code -m anthropic/claude-opus-5-5 --allow-agent-host api.anthropic.com

Note for standalone Harbor: every task.toml already sets the 6-hour agent time limit (timeout_sec = 21600 under [agent]). On top of Harbor, eval/run.py builds the task-image verifiers and raises Harbor 0.20's 15-second proxy header timeout for long rubric evaluations.

Citation

@misc{han2026ple,
  title         = {{PLE}: Physical-Engineering's Last Exam},
  author        = {Rujun Han and Yufan Zhuang and Arijit Ray and Han Yu and
                   Zoey CuiZhu and George Lee and Yifei Ming and Binz Roy and
                   Varun Sakalkar and Vishy Tirumalashetty and Burak Gokturk and
                   Tomas Pfister and Chen-Yu Lee},
  year          = {2026},
  eprint        = {2610.30537},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.30537}
}

License

The dev tasks are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Downloads last month: 1,029

Total file size:

955 MB