Datasets:
Formeo Portal Frame Dataset v2.1
Verifiable portal-frame environment: world = frame family, task = size the members, grader = per-clause EN 1993. This public sample is 180 train rows plus expert trajectories for 8 of them, under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). It is a teaser, not a training corpus. Need another structure type or engineering domain? Get in touch: info@formeo.ai.
Biaggio74/portal-frame-verifier β Try the live portal-frame grader β
Sign in with Hugging Face. Free demo limits (100 calls/day). Private keys: info@formeo.ai.
Summary
Environment + grader for steel portal frames. The grader is the product; the files demonstrate it. Design inputs, reference solutions, per-clause Eurocode verdicts, and labeled ULS/SLS outcomes.
| Role | What it is |
|---|---|
| World | Single-bay pitched-roof steel portal frame, new geometry/loads/grade per seed |
| Task | Size the members (propose column and rafter sections) |
| Grader | Deterministic per-clause EN 1993 pass/fail with utilisations (milliseconds, no compute cost per step) |
v2.1 is the open sample on this repo. 180 train rows is too small to train a model. The 8 expert trajectories are the advert for the grader: the engineer's decision sequence as structured data β section tried, governing clause, utilization, decision.
Commercial engagements cover family build, a licensed corpus, and grader access. Corpora are sized per engagement, generated fresh per licensee.
- HF dataset:
Formeo/portal-frame-dataset - Dataset version: 2.1
- Structure family: single-bay pitched-roof steel portal frame
- Public sample: 180 train samples (cc-by-nc-4.0), no val/test
- Expert trajectories: 8 frames, one row per step in
trajectories.parquet - Licensed corpora: sized per engagement, generated fresh per licensee.
- Record schema: 1.1
- Units: SI (m, kN, kNm, kg, kgCO2e)
Dataset contents
| Content | Purpose |
|---|---|
| Frame inputs (geometry / loads / grade) | World + task specification |
| Final steel sections + mass / eCOβ | Reference solution |
| Per-member EN 1993 check table | Grader verdicts (clause-level) |
| SLS deflection (actual vs L/200, H/150) | Serviceability labels for supervised learning |
failure_mode: none / uls / sls / both |
Strength vs deflection failure attribution |
status |
Design outcome including labeled non-optimal cases |
Not in this public sample: the rest of the trajectory corpus and val/test split IDs. Those stay licensed. A rate-limited live grader demo is public: https://huggingface.co/spaces/Biaggio74/portal-frame-verifier.
Acceptance rule (dataset labels and grader are identical): a design is feasible only when both ULS strength utilization and SLS deflection utilization are β€ 1.0. Governing utilization = max(ULS, SLS).
Intended use
This package is a non-commercial teaser. It is not a benchmark and not a training set: 180 train rows, with the catalog walk shown for 8 of them.
Do not report eval numbers on this sample. A clean evaluation needs tasks generated from a private holdout that has never been published.
Formeo provides a programmatic world generator and a deterministic Eurocode grader under license. Downstream systems can train or evaluate models that propose member sections; the grader returns pass/fail and per-clause diagnostics.
Parameter space (this release)
Ranges below match the factory sampler (CONTINUOUS_BOUNDS / CATEGORICALS) used to generate the data.
| Parameter | Field | Range / levels | Type |
|---|---|---|---|
| Span | span |
12β25 m | float |
| Eaves height | h_eaves |
4β10 m | float |
| Roof pitch | pitch |
3β15 deg | float |
| Frame spacing | frame_spacing |
5β8 m | float |
| Basic wind velocity | v_b0 |
22β26 m/s | float |
| Snow on ground | s_k |
0.4β0.8 kN/mΒ² | float |
| Cladding + services DL | g_k_add |
0.15β0.25 kN/mΒ² | float |
| Base fixity | base_fixity |
pinned, fixed | categorical |
| Terrain | terrain |
II, III | categorical |
| Steel grade | grade |
S275, S355 | categorical |
| Column family | column_family |
IPE, HEA | categorical |
| Rafter family | rafter_family |
IPE, HEA | categorical |
Initial section guesses are sampled within the chosen IPE/HEA family; final sections result from the sizing catalog walk (same families). Catalog upper bound in this release: IPE600 / HEA as in the Formeo materials DB.
Design status labels
| Status | Meaning |
|---|---|
optimal |
ULS and SLS both β€ 1.0; governing utilization in [0.90, 1.00] |
suboptimal |
Both pass but governing utilization < 0.90 (code-ok, overdesigned) |
infeasible_catalog_limit |
Still over-utilised; catalog could not step up |
non_converged_oscillation |
Over-utilised for other sizing reasons |
error |
Pipeline failure (not included in the published sample) |
failure_mode indicates whether the overshoot is strength (uls), deflection (sls), or both.
Check fidelity
- EN 1993 member checks: Formeo simplified implementation (not a full commercial design suite covering every clause of EN 1993-1-1).
- SLS: vertical apex Ξ΄z β€ L/200, horizontal eaves Ξ΄x β€ H/150. Actual displacements and limits are stored per sample.
- Linear 3D frame FEA; idealized single portal frame; no haunches, connections, fire, or fatigue.
- Frame-only eCOβ (default 1.5 kgCO2e/kg steel A1βA3); purlins/cladding excluded.
- Single CEN-style annex for this release.
Expert trajectories
trajectories.parquet is its own dataset config, so the Hub viewer loads it as a table. One row is one step of the engineer's catalog walk for 8 of the samples:
| Column | Meaning |
|---|---|
sample_id |
frame this step belongs to |
iter |
step number |
group |
column, roof_beam, or global |
section |
section tried |
u_gov |
governing utilization at that step |
governing_clause |
clause that governs |
limit_state |
ULS or SLS |
decision |
step_up, step_down, accept, accept_sls_floor, accept_oscillation, check, or catalog_exhausted |
The other samples ship as inputs, reference sections, and check labels only.
Package contents
samples.jsonlβ train-only sample records, without an embedded trajectorysamples.parquetβ train-only tabular rows (Hub default config)checks.parquetβ per-member / per-clause labelstrajectories.parquetβ expert-trajectory steps for 8 samples (Hub configtrajectories)docs/assets/β UI screenshot and a sample calculation report (when included)manifest.jsonβ sample size and status counts
Withheld: splits.json, val/test rows, the rest of the trajectory corpus, the generator, and the grader.
Card media lives under docs/ so Hub does not auto-detect an ImageFolder dataset from PNG directories.
Legal & Copyright
- Copyright: Β© 2026 Kisfali Engineering Srls. All rights reserved.
- Dataset public license: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) (
cc-by-nc-4.0). Attribution required. Non-commercial only. Commercial use β including training, fine-tuning, or evaluating a commercial model or product β requires a written license from Kisfali Engineering Srls. - Proprietary assets: The generation pipeline, verifier API, reasoning traces, eval holdout, full corpus, and commercial usage rights are owned exclusively by Kisfali Engineering Srls.
This license applies to v2.1. The full corpus and the rest of the expert trajectories stay with the engagement.
For commercial licensing, custom structure families, grader API access, or another engineering environment, contact Kisfali Engineering Srls. at info@formeo.ai.
Citation
@dataset{formeo_portal_frame_v2_1_2026,
author = {Kisfali Engineering Srls.},
title = {Formeo Dataset Factory: Portal Frame v2.1},
year = 2026,
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Formeo/portal-frame-dataset}
}
- Downloads last month
- 87