slugvision / README.md
kelnei's picture
Link the now-public project repo
6af475e verified
|
Raw
History Blame Contribute Delete
4.61 kB
metadata
license: gemma
task_categories:
  - image-to-text
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - synthetic
  - slug-generation
  - editorial-images
  - distillation
pretty_name: slugvision

slugvision

~32,000 image → URL-slug pairs for training small vision-language models to generate 3–5 word kebab-case permalink slugs, e.g. three-brown-teddy-bears, nursing-workforce-growth-scaling. Built as the distillation corpus for slugvision-500m and slugvision-2.2b.

The distribution mimics editorial-site image uploads (article cards): a mix of everyday photos, conceptual editorial illustrations, product shots, clinical stock imagery, and web-page screenshots. All slugs were authored by a locally hosted Gemma-family teacher (26B-class); roughly half the corpus pairs each image with the article headline it was generated for, which trains a titled mode where the slug reflects the article topic rather than just the visible content.

Slices

Slice Images Labels file Images included? Title mode
photo 20,000 photo/slugs_photos.jsonl No — COCO train2017 (see below) untitled
illustration 5,000 illustration/slugs_illustrations.jsonl Yes (FLUX.1-schnell renders) titled
product 2,500 product/slugs_products.jsonl Yes (FLUX.1-schnell renders) untitled
clinical 2,500 clinical/slugs_clinical.jsonl Yes (FLUX.1-schnell renders) half titled / half untitled
screenshot 2,000 screenshot/slugs_screenshots.jsonl Yes (rendered from bundled HTML) titled

COCO images are not redistributed — they are Flickr-sourced with mixed per-image licenses. photo/slugs_photos.jsonl keys each slug by its COCO train2017 filename (e.g. 000000000776.jpg); download the images from cocodataset.org and join on filename. Every other slice's images were generated for this dataset and ship here.

Layout

photo/slugs_photos.jsonl            labels keyed by COCO filename
illustration/
  slugs_illustrations.jsonl         labels (+ generation headline as title)
  illustration_prompts.jsonl        teacher-authored FLUX prompts (headline, scene, style)
  images/illus-NNNNNN.jpg           5,000 renders
product/   (same shape: labels, prompts, images/prod-NNNNNN.jpg)
clinical/
  ... labels, prompts, images/clin-NNNNNN.jpg
  clinical_titles_half.jsonl        which images were labeled with a title
screenshot/
  slugs_screenshots.jsonl           labels (page headline as title)
  screenshot_prompts.jsonl          page index (id, topic, headline)
  pages/shot-NNNNNN.html            2,000 teacher-authored self-contained pages
  images/shot-NNNNNN.jpg            rendered screenshots (varied viewports)
train.jsonl                         the exact 30,739-pair training split
eval.jsonl                          the frozen 1,000-image eval split (200/slice)

Label record shape:

{"image": "illus-000014.jpg", "slug": "iron-absorption-mechanism-process",
 "raw": "iron-absorption-mechanism-process",
 "title": "How Iron Absorption Works in the Human Body", "valid": true}

slug is the normalized label (lowercase kebab-case); raw is the teacher's verbatim reply; valid flags 3–5 word format compliance (97.7–99.8% per slice). title: null means the untitled prompt mode. train.jsonl / eval.jsonl records use repo-relative image paths plus a slice field — photo-slice paths point at COCO filenames you must supply yourself.

Generation pipeline

  1. Prompts/pages: the teacher authored FLUX prompts (headline + scene + style) for illustrations, products, and clinical imagery, and complete self-contained HTML pages for screenshots.
  2. Images: FLUX.1-schnell (4 steps) rendered the three generated slices; Playwright/Chromium screenshotted the HTML at weighted random viewports.
  3. Labels: the teacher labeled every image — with the source headline in context for titled slices, image-only otherwise.

All generation, training, and eval scripts live in the project repo: github.com/kelnei/slugvision.

License notes

  • Slugs, prompts, and HTML pages are outputs of a Gemma-family model — usable per the Gemma Terms of Use (hence the license tag).
  • Generated images are FLUX.1-schnell outputs (Apache-2.0 model).
  • COCO annotations/filenames referenced under CC-BY 4.0; images not included.