pretty_name: ExtFoamBench — CFD Tasks, Geometry Assets and Evaluation Metrics
language:
- en
license: other
license_name: easyfoam-benchmarks-component-licenses
license_link: LICENSE.md
size_categories:
- n<1K
tags:
- benchmark
- computational-fluid-dynamics
- openfoam
- scientific-computing
- agents
- physics
- evaluation
- reference-data
configs:
- config_name: extended
default: true
data_files:
- split: test
path: data/extended.jsonl
- config_name: basic
data_files:
- split: test
path: data/basic.jsonl
- config_name: advanced
data_files:
- split: test
path: data/advanced.jsonl
- config_name: all
data_files:
- split: test
path:
- data/extended.jsonl
- data/basic.jsonl
- data/advanced.jsonl
ExtFoamBench
English | 中文介绍
100 CFD tasks · 7 geometry assets · task-specific physical evaluation
ExtFoamBench is a benchmark for evaluating OpenFOAM agents, introduced in EasyFoam. This repository focuses on the ExtFoamBench dataset, the geometry assets needed by its tasks, and the evaluation metrics and acceptance criteria.
Core resources
| Resource | What it provides | Files |
|---|---|---|
| ExtFoamBench dataset | 100 tasks across 10 families, covering operating conditions, mesh/geometry, physics, numerical settings and multi-run workflows | Task table · Reference case inputs |
| Geometry assets | 7 ready-to-use geometry files for stirred-reactor, airfoil and vehicle cases, with source attribution and checksums | Assets · Manifest |
| Evaluation metrics and criteria | Execution, numerical and physical acceptance; per-task QoI definitions, measurement rules, reference comparisons and tolerances | Measurement specs · QoI design · Reference values and tolerances |
Each family has a baseline and nine variants: yPlusVerification, risingBubble, supersonicCrossJet, nozzleCDV, MRF_stirredReactor, ERCOFTA_pump, compressibleBubbleCollapse, aerofoilNACA0012, drivaerFastback, and riemann2D.
The tasks are developed by the EasyFoam authors. Geometry assets are redistributed from attributed OpenFOAM 13 sources; their original notices are retained.
Evaluation
A task is assessed along three dimensions: E (execution completes), N (applicable numerical checks pass), and Q (all mandatory physical/QoI criteria pass).
- ECR = tasks passing E / all assigned tasks.
- TAR = tasks passing E, N and Q together / all assigned tasks.
QoI rules specify what to measure, spatial/time support, comparison methods and acceptance tolerances. Missing outputs remain in the denominator. Diagnostic or continuous scores do not override failed mandatory criteria.
Use the dataset
Install datasets, then load the default Extended track:
from datasets import load_dataset
tasks = load_dataset("Eurus132/ExtFoamBench", "extended", split="test", revision="v1.2.0")
print(len(tasks)) # 100
print(tasks[0]["usr_requirement"])
To download geometry, criteria and reference data as well:
hf download Eurus132/ExtFoamBench --repo-type dataset --revision v1.2.0 --local-dir ExtFoamBench
The task table contains requirements and metadata. Reference case inputs and reference answers are for evaluation; keep them outside the solving agent's accessible workspace. The test label denotes benchmark tasks, not the historical skill-evolution held-out split.
Supporting resources
For broader comparison, the repository also includes FoamBench Basic (110) and Advanced (16) from upstream FoamBench. These are inherited tasks; the EasyFoam authors contribute their evaluation criteria and offline reference data. Select basic, advanced, or all (226 tasks) to browse them.
Offline QoI references and numerical summaries support evaluation across all three tracks. The optional evaluation tools bundle the scorer, scripts and schemas; extract them only when needed. The guide covers usage, data coverage and protocol details. The EasyFoam Skill library is distributed separately.
Citation and licenses
Cite EasyFoam when using ExtFoamBench or its evaluation/reference resources, and credit FoamBench when using Basic/Advanced. See component licenses for project and third-party terms.
Version 1.2.0 uses concise, consistent filenames across data, tools and documentation. Numerical reference values, acceptance rules and scoring logic are unchanged. Earlier releases remain available.