Create README.md (#1)

README.md

@@ -0,0 +1,250 @@
+---
+language:
+- en
+tags:
+- youtube
+- thumbnails
+- image-text
+- multimodal
+- text-to-image
+- image-to-text
+- captioning
+- weak-supervision
+- large-scale
+- computer-vision
+- nlp
+- vision-language
+- clip-training
+- diffusion
+- generative-models
+- image-generation
+- thumbnail-generation
+- social-media
+- content-creation
+- visual-design
+- high-contrast
+- faces
+- expressions
+- memes
+- clickbait
+- marketing
+- advertising
+- attention-modeling
+- representation-learning
+- embedding
+- retrieval
+- search
+- ranking
+- dataset-creation
+- public-data
+- self-supervised
+- weak-labels
+- noisy-labels
+- english
+- filtered
+- deduplicated
+- large-dataset
+- research
+- experimental
+- open-data
+- vision
+- multimodal-learning
+- image-dataset
+- text-dataset
+pretty_name: Youtube Thumbnails
+task_categories:
+- text-to-image
+- image-to-text
+- feature-extraction
+size_categories:
+- 100K<n<1M
+license: other
+---
+
+# YouTube Thumbnails Dataset
+
+## Dataset Details
+
+### Dataset Description
+
+This dataset contains approximately **164,000 YouTube thumbnails** paired with their corresponding video titles.
+
+The dataset was constructed by collecting public YouTube channel feeds, extracting video metadata, filtering and deduplicating entries, and downloading thumbnail images at scale.
+
+The goal of this dataset is to support research and experimentation in:
+
+- Image generation (e.g. diffusion models)
+- Multimodal learning (e.g. CLIP-style models)
+- Thumbnail generation and optimization
+- Image-text representation learning
+
+---
+
+- **Curated by:** l3af (Discord: l3afai)  
+- **Language(s):** English (filtered using language detection)  
+- **License:** Derived from publicly available YouTube data. Users are responsible for complying with YouTube's Terms of Service.  
+
+---
+
+## Dataset Sources
+
+- **Source:** Public YouTube RSS feeds (`videos.xml`)
+- **Images:** YouTube thumbnail CDN (`i.ytimg.com`)
+- **Metadata:** Video titles and IDs
+
+---
+
+## Uses
+
+### Direct Use
+
+This dataset is suitable for:
+
+- Training image generation models (especially thumbnail-style generation)
+- Training multimodal embedding models (e.g. CLIP)
+- Studying social-media visual patterns
+- Thumbnail generation or ranking systems
+
+---
+
+### Out-of-Scope Use
+
+This dataset is **not recommended for**:
+
+- High-quality caption-to-image generation (titles are not descriptive captions)
+- Tasks requiring precise semantic grounding
+- Sensitive or safety-critical applications
+
+---
+
+## Dataset Structure
+
+Each example contains:
+
+- `video_id` (string): YouTube video identifier  
+- `title` (string): Video title  
+- `image` (image): Thumbnail image  
+
+---
+
+## Dataset Creation
+
+### Curation Rationale
+
+This dataset was created to provide a large-scale collection of real-world image-text pairs with strong visual patterns, particularly useful for studying:
+
+- Attention-grabbing design
+- High-contrast visual composition
+- Social media aesthetics
+
+---
+
+### Source Data
+
+#### Data Collection and Processing
+
+The dataset was created through the following pipeline:
+
+1. Collected ~22,000 YouTube channel IDs
+2. Downloaded RSS feeds (`videos.xml`)
+3. Extracted video metadata
+4. Filtered:
+   - Removed Shorts content
+   - Removed non-English titles (via language detection)
+5. Deduplicated titles (exact + fuzzy)
+6. Downloaded thumbnail images (max resolution when available)
+7. Built dataset in multiple formats (Parquet + HF dataset)
+
+---
+
+#### Who are the source data producers?
+
+The source data was originally created by YouTube content creators across a wide range of domains, including entertainment, education, gaming, and news.
+
+---
+
+### Annotations
+
+No manual annotations were added. The dataset consists solely of:
+
+- Original thumbnails
+- Original video titles
+
+---
+
+### Personal and Sensitive Information
+
+- Some thumbnails may contain human faces or identifiable individuals
+- Titles and images may reflect biases from content creators
+- No additional personal data was intentionally collected
+
+---
+
+## Data Traceability
+
+Each entry includes a `video_id` which uniquely identifies the original YouTube video.
+
+Users can reconstruct the original source via:
+
+https://www.youtube.com/watch?v={video_id}
+
+This enables:
+- Attribution to original creators
+- Verification of data origin
+- Selective filtering or removal
+
+---
+
+## Bias, Risks, and Limitations
+
+- Strong bias toward YouTube-style content (faces, text overlays, high contrast)
+- Titles are often:
+  - Clickbait
+  - Vague
+  - Non-descriptive
+- Images frequently contain embedded text (which models struggle to generate correctly)
+- Distribution may not reflect real-world image diversity
+
+---
+
+### Recommendations
+
+- Use for **style-focused tasks**, not semantic grounding
+- Consider augmenting with caption datasets for better text alignment
+- Filter further if targeting specific domains
+
+---
+
+## Citation
+
+If you use this dataset, please cite:
+```
+l3afai. (2026). YouTube Thumbnails Dataset.
+```
+
+## Dataset Card Authors
+
+- l3af (Discord: l3afai)
+
+## License and Attribution
+
+This dataset contains images and metadata derived from publicly available YouTube content.
+
+- All rights to the original thumbnails belong to their respective creators.
+- This dataset does not claim ownership of any images.
+- Each sample includes a `video_id` which can be used to trace the original source:
+  https://www.youtube.com/watch?v={video_id}
+
+This dataset is provided for research and educational purposes only.
+
+If you are a content owner and would like your data removed, please contact the dataset maintainer.
+
+## Takedown Policy
+
+If you are a rights holder and wish to have content removed from this dataset, please contact the maintainer with the relevant `video_id`(s). The content will be removed.
+
+## Dataset Card Contact
+
+For questions or issues, contact:
+
+- Discord: l3afai