EthanGabis/ADautoGen-DS

ADautoGen-DS: Multi-Modal Product Advertisement Dataset

A synthetic dataset of 120 product advertisements with AI-generated text and images, designed for multi-modal recommendation systems research.

Dataset Overview

Dataset Structure

ADautoGen-DS/
├── id                 # Unique product identifier
├── product_name       # Short product name (avg 18.9 chars)
├── category           # Product category (5 classes)
├── tagline            # Marketing tagline (avg 53.3 chars)
├── ad_copy            # Full advertisement text (avg 120.9 chars)
├── target_audience    # Target demographic
├── source_model       # LLM that generated the text
└── image              # Product image (PIL Image)

Exploratory Data Analysis

1. Text Length Distributions

[image]

Key Insights:

• —Product Names follow a normal distribution centered at 20 characters, with most names between 14-24 characters. This indicates consistent, concise naming conventions suitable for display in UI components and search results.

• —Taglines show a bimodal distribution with peaks at ~35 and ~80 characters. The 80-character cluster suggests many products received detailed taglines, while shorter ones are more punchy marketing phrases. This variability adds realism to the dataset.

• —Ad Copy clusters heavily around 140 characters with a secondary peak near the maximum. The distribution shows the LLMs learned to produce social-media-ready copy lengths, similar to Twitter/X character limits, making this dataset practical for real-world advertising applications.

2. t-SNE: Text Embeddings by Category

[image]

Key Insights:

• —Clear Semantic Clustering: Categories form distinct regions in the embedding space. Beauty products (gray) cluster in the upper-right, while Fitness products (brown) occupy the lower-left quadrant.

• —Category Separation: Food (blue), Tech (green), and Home (cyan) products show meaningful separation, confirming that the text embeddings capture genuine semantic differences between product types.

• —Some Overlap Exists: The moderate overlap between categories (especially Food/Home and Tech/Home) reflects real-world product similarities - for example, kitchen appliances could logically belong to either Food or Home categories.

• —Embedding Quality: The MiniLM text encoder successfully maps product descriptions to semantically meaningful vectors, validating its use for text-based product search.

3. t-SNE: Text Embeddings by Source Model

[image]

Key Insights:

• —No Model Clustering = Success: Points from Phi-2 (circles), Qwen-1.8B (squares), and TinyLlama (triangles) are uniformly distributed throughout the embedding space with no visible clustering by source model.

• —Text Contamination Avoided: This visualization proves the multi-model generation strategy worked. If a single model had generated all text, we would see systematic biases in writing style that would cluster together.

• —Genuine Diversity Achieved: The random interleaving of all three marker types across all regions confirms that text diversity comes from semantic content (category, product type) rather than model-specific patterns.

• —Validates Multi-LLM Approach: Using 3 different small language models (instead of one large model) successfully prevented the "AI slop" homogeneity problem common in synthetic datasets.

4. Pairwise Text Similarity Distribution

[image]

Key Insights:

• —Mean Similarity: 0.249 - This relatively low mean indicates high text diversity. Products are semantically distinct from each other on average.

• —Distribution Shape: The right-skewed distribution (peak at 0.15-0.20) shows most product pairs have low similarity, with a long tail extending to 0.8+ for genuinely similar products.

• —High Similarity Tail: The ~5% of pairs with similarity >0.6 likely represent products in the same category or serving similar functions - this is expected and desirable behavior.

• —No Duplicates: Virtually no pairs show similarity >0.9, confirming no duplicate or near-duplicate products exist in the dataset.

• —Recommendation Viability: The diversity distribution is ideal for recommendation systems - enough similarity to find related products, but enough diversity that recommendations are meaningful.

5. RGB Channel Distribution

[image]

Key Insights:

• —Balanced Color Palette: Average RGB values (133.3, 127.9, 128.7) are nearly identical across channels, indicating the image dataset has no systematic color bias.

• —Mid-tone Dominance: All three channels peak around 120-140 intensity, suggesting most images use professional, neutral backgrounds (whites, grays) typical of product photography.

• —Blue Channel Variance: The blue channel shows slightly higher variance (extending to 180+), likely from tech product images featuring blue accent colors and LED indicators.

• —Professional Aesthetic: The tight clustering around mid-tones and absence of extreme values (very dark/bright) confirms Stable Diffusion generated clean, studio-style product images.

• —Consistent Lighting: The similar distributions across R, G, B indicate consistent white-balanced lighting across all generated images.

6. t-SNE: Image Embeddings by Category

[image]

Key Insights:

• —Strong Visual Clustering: Image embeddings show even clearer category separation than text embeddings. Tech products (green) form a tight cluster in the upper-left, completely separated from Fitness (brown) in the lower-right.

• —Visual Distinctiveness: Categories that look different cluster apart. Beauty products (gray, upper region) and Home products (cyan, left side) occupy distinct visual spaces.

• —SigLIP's Power: The 768-dimensional SigLIP embeddings effectively capture visual product characteristics - shape, color, texture - that distinguish categories.

• —Fitness Isolation: Fitness products form the most isolated cluster (bottom-right), likely because gym equipment, resistance bands, and supplements have distinctive visual features unlike other categories.

• —Image-Based Search Viability: This clear clustering enables effective "find similar products" functionality based purely on visual appearance.

7. Text vs Image Embedding Comparison

[image]

Key Insights:

• —Complementary Modalities: Text (MiniLM, left) and Image (SigLIP, right) embeddings show different clustering patterns, confirming they capture complementary information.

• —Tighter Image Clusters: Image embeddings show more compact, well-defined category clusters compared to the more diffuse text clusters. Visual features are more consistently distinctive.

• —Different Category Relationships: In text space, Food and Home overlap significantly. In image space, Tech and Home have more overlap. This asymmetry is why multi-modal embeddings outperform single-modality approaches.

• —Multi-Modal Justification: A product might have similar description to one category but look similar to another. Combining both modalities captures this nuance for better recommendations.

8. Text vs Image Similarity Distributions

[image]

Key Insights:

• —Text More Diverse (Mean: 0.249): Text descriptions show lower average pairwise similarity, indicating greater linguistic diversity in how products are described.

• —Images More Uniform (Mean: 0.567): Image embeddings show higher average similarity (0.567 vs 0.249), reflecting the consistent studio-style product photography generated by Stable Diffusion.

• —Different Distributions: Text similarities are right-skewed (most pairs dissimilar), while image similarities are more normally distributed around the mean.

• —Single Generator Effect: The higher image uniformity is expected since all images came from one Stable Diffusion model, while text came from three different LLMs.

• —Implications for Search: Text search will find more distinct results; image search will find more visually similar products. The combined embedding balances both behaviors.

9. EDA Summary Statistics

[image]

Key Findings Summary:

1. 1.Multi-Model Strategy Worked: Using 3 different LLMs (Phi-2, Qwen-1.8B, TinyLlama) successfully reduced text contamination - no model-based clustering detected.

1. 1.Semantic Meaning Captured: t-SNE visualizations show category clustering in both text and image spaces, confirming embeddings capture product semantics.

1. 1.Genuine Diversity Achieved: Low cross-model similarity (0.239) vs within-model similarity (0.270) proves different models produce different writing styles.

1. 1.Visual Consistency: Images are more uniform due to single Stable Diffusion generator, but still show clear category separation.

Usage

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("EthanGabis/ADautoGen-DS", split="train")
print(f"Loaded {len(ds)} products")

Accessing Data

# Get a sample
sample = ds[0]
print(f"Product: {sample['product_name']}")
print(f"Category: {sample['category']}")
print(f"Generated by: {sample['source_model']}")

# Display image
sample['image'].show()

Loading Pre-computed Embeddings

import numpy as np

# Combined text+image embeddings (768-dim SigLIP)
emb_combined = np.load("emb_combined_siglip.npy")
print(f"Shape: {emb_combined.shape}")  # (120, 768)

Live Demo

Try the multi-modal recommendation system: HuggingFace Space

Citation

@dataset{adautogen_ds_2026,
  title={ADautoGen-DS: Multi-Modal Product Advertisement Dataset},
  author={Ethan Gabis},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/EthanGabis/ADautoGen-DS}
}

License

MIT License - Free for academic and commercial use.