Datasetpublic

leixiang25/24679-hw1-image-register

24-679 HW1 (Fall 2026): Business Message Register Images leixiang25/24679-hw1-image-register Digitally rendered screenshot-style images of short, fictional business messages, labeled by register. 1 = formal (high-context business register); 0 = casual (low-context register). Created by Lei Xiang for 24-679 Homework 1 at Carnegie Mellon University. The dataset is related to Context, a cross-cultural deal interpreter for Western operators working with Japanese and Chinese… See the full description on the dataset page: https://huggingface.co/datasets/leixiang25/24679-hw1-image-register.

sourceHugging Faceupdated 9d agoView on Hugging Face

0likes59downloads

Dataset Card

24-679 HW1 (Fall 2026): Business Message Register Images

leixiang25/24679-hw1-image-register

Digitally rendered screenshot-style images of short, fictional business messages, labeled by register. 1 = formal (high-context business register); 0 = casual (low-context register). Created by Lei Xiang for 24-679 Homework 1 at Carnegie Mellon University. The dataset is related to Context, a cross-cultural deal interpreter for Western operators working with Japanese and Chinese counterparts, where register is one of the first signals read from an incoming message.

Source and task

The 24 messages are fictional, drafted with AI assistance (Claude) and reviewed and labeled by the author. They contain no names, faces, or personal information; roles ("Counterpart") replace names. Each message is rendered in two styles (an email panel and a chat bubble), giving 48 original images. Labels are assigned by the author at writing time from the register of the wording.

Fields

Field	Meaning and modeling role
`image`	224 x 224 RGB working copy, padded with gray (128, 128, 128); the model input.
`label`	Binary target: 1 = formal, 0 = casual.
`label_name`	Readable target; exclude from inputs.
`message_id`	Underlying message key shared by both renders; grouping/provenance only.
`render_style`	`email` or `chat`; provenance only.
`source_id`, `parent_id`	Unique example key and its original image key; provenance only.
`augmentation`, `is_augmented`	Method identifier and synthetic flag; provenance only.

Splits

Split by message_id (stratified by label) before any augmentation, so both renders of a message stay in one partition. Requested holdout 30%, half of it to test. Seed 24679 for the first split and 24680 for the holdout split. Validation and test contain only unaugmented working copies.

Split	Original rows	Synthetic rows	Total rows
train	32	384	416
validation	8	0	8
test	8	0	8

Split	Label	Rows
train	0 (casual)	208
train	1 (formal)	208
validation	0 (casual)	4
validation	1 (formal)	4
test	0 (casual)	4
test	1 (formal)	4

Augmentation and preprocessing

Every image is resized to fit 224 x 224 with aspect ratio preserved and centered on a gray canvas. Each training original contributes 3 independent variants from each of four methods, each applied to the prepared parent (never stacked): brightness factor 0.55 to 0.80 or 1.20 to 1.45; rotation 3 to 8 degrees either direction; contrast factor 0.55 to 0.80 or 1.25 to 1.70; Gaussian blur radius 0.5 to 1.1 px. The ranges are milder than the classroom bin dataset on purpose: the label evidence is small printed text, and stronger rotation or blur would make it unreadable.

Training method	Stored rows
`mild_brightness`	96
`mild_contrast`	96
`mild_gaussian_blur`	96
`none`	32
`small_rotation`	96

Intended use and limitations

Teaching dataset for image preparation, augmentation auditing, and transfer learning. It is small, synthetic, and stylistically narrow (two templates, one font family), so a model can shortcut on layout rather than reading. Register is a simplification of real cross-cultural signals. Synthetic variants do not add independent messages. Report per-class counts with accuracy or F1.

Privacy and licensing

No real people, faces, or identifying information appear. Content is original fictional text drafted with AI assistance and reviewed by the author. No license is assigned by this card.

Load

python

from datasets import load_dataset
ds = load_dataset("leixiang25/24679-hw1-image-register")