ishaanamahajan/hw1-paper-triage-multimodal
HW1 Multimodal Paper Triage Purpose This dataset supports a classroom exercise in assembling and augmenting multimodal data for a personalized research-paper triage system. Composition and splits The dataset begins with 100 original multimodal samples. The original samples were split before augmentation using a fixed random seed and stratification by the binary target. train: 10,070 samples consisting of 70 training originals and 10,000 augmented… See the full description on the dataset page: https://huggingface.co/datasets/ishaanamahajan/hw1-paper-triage-multimodal.
HW1 Multimodal Paper Triage
Purpose
This dataset supports a classroom exercise in assembling and augmenting multimodal data for a personalized research-paper triage system.
Composition and splits
The dataset begins with 100 original multimodal samples. The original samples were split before augmentation using a fixed random seed and stratification by the binary target.
train: 10,070 samples consisting of 70 training originals and 10,000 augmented versions generated only from those training originals.validation: 15 unaugmented original samples.test: 15 unaugmented original samples.- Each sample contains seven tabular features, a short text descriptor, a 224×224 image, and a binary target.
Original sample creation
The text descriptors are hypothetical abstract-like descriptions organized into twenty topic groups. Each description is paired with a digitally created neutral paper-preview card. The cards display title and metadata but never display the target. These examples are not abstracts or metadata from specific published papers.
The seven primary tabular features are publication_year, estimated_pages, title_word_count, text_character_count, keyword_count, interest_overlap_count, and method_word_count. Additional categorical metadata records topic, method, goal, evaluation setting, and focus.
Target
should_read = 1 indicates an immediate match to a fixed interest profile covering robot learning, manipulation, diffusion and generalist robot policies, world models, test-time adaptation, safety and recoverability, human recourse, legged control, and efficient onboard deployment. should_read = 0 indicates that the description falls outside that immediate profile. The label reflects one reader profile, not scientific quality.
Augmentation and preprocessing
Only the 70 training originals were augmented. Exactly 10,000 augmented training samples were generated, with each training source producing either 142 or 143 variants. No augmentation was applied to the validation or test samples.
Text augmentation uses controlled templates that preserve the source topic, method, goal, evaluation, and focus. Image augmentation applies rotations of at most four degrees, translations of at most three pixels, mild brightness and contrast adjustments, and occasional light blur. These transformations preserve the target because they do not change the represented paper topic or its relevance to the fixed interest profile.
Intended use and limitations
This dataset is intended for coursework involving multimodal dataset handling and simple relevance prediction. It should not be used to infer universal preferences, paper quality, or citation impact. The samples are hypothetical, the labels reflect one narrow profile, and repeated templates or card layouts may introduce shortcuts that a model can exploit.
Ethics and privacy
The dataset contains no people, faces, private information, or sensitive imagery. It should not be presented as a corpus of real scientific abstracts.
License
The dataset and code are released under CC BY 4.0.
Collaboration and GenAI disclosure
The paper-triage project topic was discussed with the student's project partner, Sri. OpenAI Codex assisted with the hypothetical descriptions, notebook code, digital-card generator, augmentation logic, and documentation.