---
license: other
license_name: internal-research-only
license_link: LICENSE.md
pretty_name: FrontierBench CAD Authoring Seeds
language:
  - en
tags:
  - cad
  - engineering-drawing
  - gd-and-t
  - synthetic-data
extra_gated_prompt: >-
  This dataset is released for internal research review only. Access is granted
  per request. By requesting access you agree not to redistribute the contents
  and not to use them to train publicly released models. Note that `source-audit/`
  holds upstream third-party CAD files carrying enterprise authorship metadata,
  included by operator decision for audit purposes; redistribution authority for
  those files is disputed and is not granted to you by this dataset.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree not to redistribute or attempt source re-identification: checkbox
configs:
  - config_name: profile_metrics
    data_files: metrics/profile_metrics.jsonl
---

# FrontierBench CAD Authoring Seeds

Measured design-complexity profiles and clean-room CAD generators used to author
parametric-rebuild benchmark tasks. Everything under `profiles/` and
`clean-room/` is either a statistic measured from source material or first-party
generator code. **`source-audit/` is the exception: it holds unmodified upstream
third-party CAD files**, added by operator decision — see `RIGHTS.md` before
using them for anything.

## Contents

| Path | Count | What it is |
| --- | ---: | --- |
| `profiles/PGS-*.json` | 15 | Per-part design-complexity profiles: dimension counts by kind and qualifier, tolerance deviation magnitudes, nominal ranges, GD&T type histograms, datum-frame shapes, per-measure authoring units. Schema `2.0`. No geometry, no coordinates, no source identifiers. |
| `clean-room/*/truth.json` | 11 | Frozen design definitions for first-party parametric families: parameter space or discrete variation lists, derived quantities, named constraints, drawing contract, coordinate frame, units. |
| `clean-room/*/generate_assets.py` | 11 | First-party deterministic generators. Seeded sampling, no network access, no CAD addons. |
| `metrics/profile_metrics.jsonl` | 598 rows | The same profile numbers in long format, one row per measurement: `seed_id`, `metric`, `key`, `value`. Derived from `profiles/` by the builder; this is what the dataset viewer reads. |
| `source-audit/` | 3 | Unmodified upstream material for one third-party part: a STEP solid carrying embedded PMI, its drawing PDF, and the upstream collection's licence text. Not first-party, not clean-room, not produced by the builder. Carries the upstream part number, enterprise authorship and an individual's account name in its embedded metadata. |

## How to use

Access is gated: request it on the dataset page, then authenticate locally with
`hf auth login` or by exporting `HF_TOKEN`. Every snippet below was run against
this repository as published, and the outputs shown are the real ones.

**The numbers, as a table.** 598 rows, four columns, one row per measurement.

```python
from datasets import load_dataset

metrics = load_dataset(
    "SueMintony/frontierbench-cad-authoring-seeds", "profile_metrics", split="train"
)
metrics[0]
# {'key': 'inch', 'metric': 'authoring_units', 'seed_id': 'PGS-01', 'value': 1.0}
```

**One row per part, if you want a feature matrix.** Fold `key` into the column
name and pivot. The 598 rows become 15 x 96 -- 41.5% filled, because a bucket
that a part never uses has no row. Those gaps are absences, not zeros: `fillna(0)`
is right for histogram counts and wrong for magnitudes and ranges, so choose per
metric rather than globally.

```python
df = metrics.to_pandas()
df["column"] = df["metric"] + df["key"].map(lambda k: f"[{k}]" if k else "")
wide = df.pivot(index="seed_id", columns="column", values="value")
wide.loc["PGS-01", "dimension_profile.by_kind[size]"]   # 9.0
```

`(seed_id, metric, key)` is unique, so the pivot needs no aggregation.

**The profile documents.** `profiles/*.json` is the source of truth; the table is
derived from it by the builder. Fetch just those:

```python
from huggingface_hub import snapshot_download

local = snapshot_download(
    "SueMintony/frontierbench-cad-authoring-seeds",
    repo_type="dataset",
    allow_patterns=["profiles/*.json"],
)
```

Or take the whole bundle without the third-party material:

```bash
hf download SueMintony/frontierbench-cad-authoring-seeds --type dataset \
  --exclude "source-audit/*" --local-dir ./seeds
```

**Ids run `PGS-01` to `PGS-14` and then `PGS-16`** -- 15 profiles, and the
numbering skips one. `PGS-15` was rejected during measurement, not lost: its
source tagged every dimension in inch and every geometric-tolerance magnitude in
millimetre, two disjoint unit sets, which no real drawing carries. It was read as
an upstream export defect and dropped before any measurement was kept, so no
profile exists to publish under that id, and the id stays a placeholder rather
than being reused so the others are not renumbered. A loader that assumes a dense
range will go looking for a file that is not here.

**Before computing on any magnitude, read the units caveat below.** Fields already
normalised to millimetres carry a `_mm` suffix; `authoring_units` says what the
part was drawn in, and one part can legitimately mix the two.

**`source-audit/` is not analysable data.** It is unmodified upstream third-party
CAD, kept for audit and excluded from the snippets above on purpose. Read
`RIGHTS.md` before touching it, and `EXCLUDED.md` for what a rebuild does and
does not reproduce.

## Geometry families

`flanged-bushing` (4 designs), `flanged-bearing-support-bracket` (1),
`finned-electronics-mounting-plate` (3 design ids across 4 directories),
`rectangular-equipment-platform-housing` (1), sprocket (1).

One further design is still withheld in full -- its `truth.json` and generator
are not here; see `EXCLUDED.md`. The upstream material that design was measured
from is, however, now published under `source-audit/`, so the withholding no
longer protects that design from retrieval.

## What is rewritten before publishing

Selection is an allowlist and the scan is a refusal gate, so nothing here was
scrubbed into shape. Two rewrites are declared: each `source_taxonomy` keeps its
"nothing was copied" assertions but not the source's identity, and four
generators had an absolute rasteriser path replaced by the bare command, which
resolves through `PATH`. Generator logic is unchanged, and `EXCLUDED.md` lists
both alongside everything withheld outright. `source-audit/` bypasses this
pipeline entirely: those bytes are the upstream originals, unscanned and
unredacted, and rebuilding the bundle will not reproduce them.

## Why the viewer reads a second copy

`profiles/*.json` cannot be loaded as a table. Five of their fields --
`by_kind`, `by_qualifier`, `note_types`, `by_type`, `datum_frames` -- are sparse
histograms whose buckets differ per part: 9 distinct key sets for `by_qualifier`
across 15 files, 14 for `datum_frames` with a 42-bucket union. A table column has
one type, so a loader infers a struct from the first file and then fails to cast
the next one, taking the whole split with it. Widening every file to the union of
buckets would only move the problem: the buckets are data-dependent labels, so the
next part measured brings a new one and breaks it again.

`metrics/profile_metrics.jsonl` is therefore the viewer's copy, in long format
where a new bucket is another row rather than another column. Three string columns
and one float, always. A named scalar carries the dotted path with an empty key
(`dimension_profile.count`), a histogram bucket carries its label in the key
(`gdt_profile.datum_frames` / `A|B|C`), a list element carries its index, and a
string value moves into the key with `1.0`. Read `profiles/` for the documents;
read this for the numbers.

## Units caveat

Length units are recorded **per measure**, not per file. In the source STEP
files the header SI unit is only the SI base; each measure carries its own unit
reference, and a single part can legitimately mix inch dimensions with
millimetre GD&T magnitudes. `profiles/*.json` therefore carries an
`authoring_units` list, and every magnitude field is normalised to millimetres
with a `_mm` suffix. Reading a file header instead of the per-measure unit is a
silent 25.4x error.

## Known gap

Across all 15 profiles the dimension mix is `size`-dominated
(`location:size` roughly 40:198) and basic/TED dimensions are nearly absent
(6 total). The common real-world pattern of "basic dimensions locate the
features, a POSITION tolerance references datum frame A|B|C" is therefore
under-represented. Do not treat this pool as a representative sample of
production MBD practice.