CoolFace
Datasetpublic

eandujar/2026-24679-HW1-Multimodal-Original

Straight-member torque: image and structured statics data eandujar/2026-24679-HW1-Multimodal-Original 100 synthetic planar-statics cases containing a rendered diagram and structured numerical/categorical features describing the member geometry, supports, and applied loads. The prediction task has two outputs: Torque direction — clockwise or counterclockwise. Torque magnitude — absolute moment about the pin in N m. This therefore supports both classification and regression… See the full description on the dataset page: https://huggingface.co/datasets/eandujar/2026-24679-HW1-Multimodal-Original.

sourceHugging Faceupdated 11d agoView on Hugging Face
0likes86downloads
Dataset Card

Straight-member torque: image and structured statics data

eandujar/2026-24679-HW1-Multimodal-Original

100 synthetic planar-statics cases containing a rendered diagram and structured numerical/categorical features describing the member geometry, supports, and applied loads.

The prediction task has two outputs:

  1. 1.Torque direction — clockwise or counterclockwise.
  2. 2.Torque magnitude — absolute moment about the pin in N m.

This therefore supports both classification and regression experiments.

Source and task

The source DataFrame and prepared image list are joined by image filename. Each row represents one straight-member statics problem with:

  • —one pin support,
  • —one x-constrained/y-free roller support,
  • —two interior applied loads,
  • —geometric coordinates,
  • —force components,
  • —and a rendered diagram.

The signed ground-truth torque is stored in applied_moment_about_pin_Nm.

The two prediction targets are derived from that quantity:

  • —negative torque -> clockwise,
  • —positive torque -> counterclockwise,
  • —torque_magnitude_Nm = abs(applied_moment_about_pin_Nm).

The classification encoding is:

  • —0 = clockwise
  • —1 = counterclockwise

applied_moment_about_pin_Nm, torque_magnitude_Nm, label, and label_name are targets and must not be supplied as model inputs.

Fields

FieldMeaning and modeling role
imageInput: prepared diagram image with embedded image bytes.
pin_x, pin_yInput: pin-support coordinates in meters.
roller_x, roller_yInput: roller-support coordinates in meters.
load1_tInput: normalized location parameter for load 1 along the member.
load1_modeInput: active components of load 1: x, y, or xy.
load1_x, load1_yInput: load-1 application coordinates in meters.
load1_Fx, load1_FyInput: load-1 force components in newtons.
load2_tInput: normalized location parameter for load 2 along the member.
load2_modeInput: active components of load 2: x, y, or xy.
load2_x, load2_yInput: load-2 application coordinates in meters.
load2_Fx, load2_FyInput: load-2 force components in newtons.
pin_FxStructured support-reaction feature. See leakage note below.
labelClassification target: 0 = clockwise, 1 = counterclockwise.
label_nameHuman-readable direction target. Exclude from model inputs.
torque_magnitude_NmRegression target: nonnegative magnitude of torque about the pin.
applied_moment_about_pin_NmSigned ground-truth torque. Exclude from model inputs.
image_fileSource image filename / case identity. Exclude from model inputs.
source_idStable source-case identifier. Exclude from model inputs.
parent_idParent case identifier used to track derived/augmented examples.
augmentationAugmentation type; none for original examples.
is_augmentedWhether the example is an augmented derivative.
indexOriginal source-row identifier; provenance only.

Packaged splits

SplitCasesClockwiseCounterclockwise
train703634
validation15411
test1587

Holdout fraction: 30%; test receives 50% of holdout. The resulting target proportions are 70% training, 15% validation, 15% testing. Splits are stratified by class using seeds 24679 and 24680; rounding can affect proportions in small datasets. Images and descriptions stay together. There are 100 distinct cases. All cases are synthetic simulations; no augmented variants are added. Duplicate IDs and image hashes are rejected. Reuse these exact partitions for model comparisons.

Generation and units

Coordinates and distances use meters, forces use newtons, and torque uses N m. Positive x points right, positive y points up. Two applied loads may have x, y, or both components. The supports have distinct heights and are ideal bilateral constraints. Self-weight and applied couples are excluded. Descriptions round values to six decimals. Images contain no text or axes; the framing preserves geometry aspect ratio and leaves a small edge margin.

Intended use and limitations

Use for educational image-only, text-only, or combined image/text classification. The description contains enough numerical information to solve the task without the image; compare unimodal baselines to measure any multimodal benefit. Thin lines and overlapping loads can be difficult to distinguish at 224 x 224. This small synthetic dataset is not a real-world structural safety benchmark. Select settings on validation; use the test split for final evaluation.

Licensing

No license is assigned by this preparation code.

Load

python
from datasets import load_dataset
ds = load_dataset("eandujar/2026-24679-HW1-Multimodal-Original")
# Train with ds["train"], choose settings with ds["validation"], then score ds["test"].