Datasets:

Logolabs
/

LogoBrief-10K

license: other
license_name: logobrief-10k-research-license
license_link: LICENSE
pretty_name: LogoBrief-10K
gated: true
extra_gated_heading: Request access to LogoBrief-10K
extra_gated_description: >-
  This dataset contains third-party trademarks collected from the public web.
  Access is granted individually, for non-commercial research only, and is
  reviewed manually. Please complete every field — incomplete or generic
  requests are declined.
extra_gated_button_content: Accept terms and request access
extra_gated_prompt: >-
  Before requesting access, please read the LICENSE, NOTICE-TRADEMARKS and
  LEGAL-BASIS files in this repository. Every logo here is a trademark of its
  owner; LogoLabs grants no rights in any mark. Rights holders may have their
  marks removed on request, with no justification required, and you must honour
  those removals in your own copy.
extra_gated_fields:
  Full name: text
  Institution or organisation: text
  Institutional or organisational email: text
  Country: country
  Personal or organisational webpage (for verification): text
  Primary purpose:
    type: select
    options:
      - Academic research
      - Non-commercial independent research
      - Teaching or coursework
      - label: Other (describe below)
        value: other
  Describe your intended use in 2-3 specific sentences (generic answers such as "research" or "AI" are declined): text
  Do you intend to train a generative model that produces logos or brand marks?:
    type: select
    options:
      - 'No'
      - Yes - research only, outputs not published or deployed
      - Yes - outputs will be published or deployed
  I confirm my use is NON-COMMERCIAL, and that neither I nor my organisation will use this dataset or anything derived from it in a commercial product or service: checkbox
  I will NOT use this dataset to produce marks that are identical or confusingly similar to any included brand, or that could be passed off as originating from or endorsed by it: checkbox
  I will NOT redistribute this dataset or any substantial part of it, including on other dataset platforms: checkbox
  I will monitor the published removals list and delete withdrawn entries from my copy within 30 days of notification: checkbox
  I have read the Licence, the Trademark Notice and the Legal Basis statement, and I understand the images are third-party trademarks in which no rights are granted to me: checkbox
task_categories:
  - image-classification
  - image-to-text
  - text-to-image
  - feature-extraction
  - image-feature-extraction
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - logo
  - brand-identity
  - graphic-design
  - visual-design
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/**
      - split: validation
        path: data/validation/**
      - split: test
        path: data/test/**

LogoBrief-10K

10,000 brand logos, each with its original SVG and a clean raster render, plus design annotations: open-vocabulary style tags, a one-sentence motif description, a full generated design brief, and text-region bounding boxes. Domain sampling is stratified by web-popularity rank rather than selected for recognizable brands. Every included domain was checked for AI-training opt-out signals immediately before publication (see Opt-out audit evidence).

Video overview

Watch the three-minute dataset overview · Transcript

_{28 rows from the release, one per category, selected at random. Plain-text wordmarks were excluded from this sample only, for visual variety.}

10,000 logos · 87 categories | 11.3% top-decile share (was 61%) | 26.31% of the draw excluded on re-audit | 100% domains re-verified pre-publish

Comparison to existing datasets

Dataset	Size	Images	Annotation
FlickrLogos-32	8K	in-the-wild photos	detection boxes
Logo-2K+	167K	in-the-wild photos	brand class
LogoDet-3K	158K	in-the-wild photos	detection boxes
OpenLogo	27K	in-the-wild photos	detection boxes
LLD	600K	32–64px thumbnails	none
LogoBrief-10K	10K	original SVGs + clean renders	design taxonomy · text geometry

The comparison datasets provide detection-oriented annotation: a bounding box locating the mark, or a brand-class label. None annotate the design itself.

Two annotation types not present in the comparison datasets:

Text-region geometry. A bounding box for every text region, in source-pixel coordinates. Usable as ground truth for whether a generative model places lettering correctly.

Design-taxonomy tags (~39 per logo): lockup, typography class, stroke contrast, terminal shape, palette structure, shape language, era. Not a brand-identity label.

Provenance and access conditions

The images are third-party trademarks. They were obtained by automated retrieval from publicly accessible web pages, in accordance with the Robots Exclusion Protocol (robots.txt) as published by each host at the time of retrieval. No permission was sought from or granted by any rights holder. LogoLabs claims no rights in any mark and grants none.

This statement covers the Robots Exclusion Protocol only. It is not a representation regarding any other means by which rights may have been reserved.

Every domain was re-checked, live, immediately before publication — robots.txt, the W3C TDM Reservation Protocol, ai.txt, llms.txt, response headers, and homepage metadata — independently of the original crawl date. A domain showing a reservation signal, or one that could not be confirmed clear, was excluded and replaced, regardless of whether the reserving text named this collector specifically. A second, model-assisted pass separately reviewed prose the pattern check does not parse for meaning — robots.txt comments and the free text of ai.txt/llms.txt — and excluded and replaced 226 further domains found there after initial publication. Full check output, including every excluded domain and its stated reason, is published; see Opt-out audit evidence.

Access is non-commercial research only, granted individually, reviewed manually.

Any rights holder may request removal of their mark, without justification. Holders of existing copies must apply removals within 30 days of notification.

Full text: LICENSE · NOTICE-TRADEMARKS.md · LEGAL-BASIS.md · TAKEDOWN.md

Sampling methodology

Most public logo corpora are constructed from top-ranked domain lists or annotated for recognizable brands first, which concentrates rows at the high-popularity end and leaves the low-popularity end sparse.

A naive intersection of annotated logos placed 61% of rows in the top decile of Tranco popularity, against 50% for the full 114K corpus. The sampling procedure instead draws from a larger pool (~88K logos), divided into 20 five-point Tranco-centile bins plus an "unranked" stratum, and allocates rows by water-filling toward an equal count per bin: bins with sufficient supply are capped at a shared ceiling; bins below the ceiling contribute their full available count.

This draw placed 13 of 21 bins at exactly 5.66% of rows; top-decile share fell to 11.3%. This is not the distribution in the shipped data — the opt-out audit changed it further.

The opt-out audit excluded 26.3% of this draw. Every domain required a live reservation check before publication, applied independently of popularity. An uneven exclusion rate across bins would have reintroduced a popularity-correlated bias, so each loss was backfilled from its own bin: fresh candidates were drawn and re-audited until every bin returned to its pre-audit count.

Eight of the 21 bins, the lowest-popularity ones, had already exhausted their eligible pool in the first draw and had no remaining supply to backfill from. Rather than under-deliver in exactly those eight bins, their losses were reassigned to the nearest bin with spare supply. This is a disclosed exception, confined to those eight bins; every other bin was backfilled from itself. Result: 10,000 rows, all 21 targets met exactly, with limited cross-bin mixing concentrated at the low-popularity end.

Procedure, in order: select_manifest.py (stratified draw) → optout_audit.py (reservation check) → select_backfill.py (replacement, with spillover) → optout_merge.py (assembly) → fix_blank_images.py (render QC). Deterministic and reproducible given the same corpus snapshot and audit results.

Access is gated. The following shows three representative rows.

data/
  train/       metadata.jsonl (9,006 rows)  images/<logo_id>.png  embeddings_dinov3_vitl16.npy
  validation/  metadata.jsonl (497 rows)    images/<logo_id>.png  embeddings_dinov3_vitl16.npy
  test/        metadata.jsonl (497 rows)    images/<logo_id>.png  embeddings_dinov3_vitl16.npy
vectors/
  train/       <logo_id>.svg (9,006 original source vectors)
  validation/  <logo_id>.svg (497 original source vectors)
  test/        <logo_id>.svg (497 original source vectors)
  optout_audit_evidence.jsonl               10,000 rows -- one per domain, global, see below
  optout_rejected_evidence.jsonl            2,772 rows -- every domain excluded, and why, see below

One render per domain. Each domain is assigned to exactly one split; no brand appears across train/validation/test, and no two rows in the release are variants of the same brand.

from datasets import load_dataset
ds = load_dataset("Logolabs/LogoBrief-10K")   # -> DatasetDict with train/validation/test

Do not load this with load_dataset("imagefolder", data_dir="data"). 18 of the 10,000 logo_id values contain "test" as a substring of the brand name (e.g. latestpokerbonuses_com_1), and 4 contain "train" (e.g. bundletraining_com_1). Against a flat data/ directory, Hugging Face's filename-based split detection matches these substrings and misassigns the entire dataset, dropping every metadata column. The three-subfolder layout above, together with this card's YAML configs, avoids that failure mode regardless of what a given logo_id contains. When loading a local clone directly, use the explicit form:

ds = load_dataset("imagefolder", data_files={
    "train": "data/train/**", "validation": "data/validation/**", "test": "data/test/**"})

Splits

90 / 5 / 5 — train 9,006, validation 497, test 497 — stratified across the same 21 popularity bins used for sampling. A random split would have reintroduced the fame-correlated bias the sampling procedure was designed to remove, at the split boundary rather than the dataset as a whole. Reproducible via select_splits.py (seeded). Counts shift slightly (from an earlier 9,002/499/499) each time the underlying 10,000-domain set changes composition, as it did during the second-pass opt-out review below — the seeded stratified draw is deterministic given its input, not given a fixed split size.

Identity

Field	Type	Notes
`logo_id`	string	stable key, also the image filename
`svg_path`	string	repository-relative path to the exact original SVG: `vectors/<split>/<logo_id>.svg`
`domain` · `brand_name` · `category`	string	87 categories
`image_width` · `image_height`	int	source-pixel dimensions; `text_boxes` are in this coordinate space

Design annotation

Field	Type	Coverage
`style_tags`	list[str]	100% — open vocabulary, ~39 tags/logo
`motif`	string	100% — one-sentence description of the mark
`caption`	string	100% — full generated design brief (typography, colour, composition)

Text geometry

Field	Type	Notes
`text_boxes`	list[[x0,y0,x1,y1]] or null	`null` = not annotated; `[]` = annotated, no text present. These are not equivalent.
`n_text_boxes`	int or null	null iff `text_boxes` is null

Aesthetic sliders

Range 0–10, LLM-rated, partial coverage: slider_creativity · slider_complexity · slider_playfulness · slider_modernity · slider_boldness · slider_colorfulness · slider_organicness · slider_luxury

DINOv3 embeddings

<split>/embeddings_dinov3_vitl16.npy — pooled ViT-L/16 embeddings (1024-d, unit-normalized, float16), one per row, ordered identically to that split's metadata.jsonl. Alignment is by row position, not by key; a train-split embedding index does not correspond to any validation-split row.

Pooled representation only: a single 1024-value vector per logo, approximately 590:1 compressed relative to the source image, with no decoder available to reconstruct pixel content. Patch-level features are a distinct, higher-fidelity representation and are not published in this release.

Coverage is non-uniform across popularity and non-monotonic. caption is 100% by construction. The fields below originate from lower-coverage annotation runs that did not reach the full corpus:

stratum n sliders text_boxes

unranked 571 45.9% 58.0%

low (0–50) 3,168 18.1% 34.2%

mid (50–85) 4,563 23.9% 37.2%

high (85–100) 1,698 48.2% 52.1%

Coverage is highest at both extremes and lowest in the middle. The images and the 100%-coverage fields are balanced by construction; the sparse fields are not. A complete-case analysis restricted to sliders or text_boxes will reintroduce the popularity bias the sampling procedure was designed to remove.

stratum	n	sliders	text_boxes
unranked	571	45.9%	58.0%
low (0–50)	3,168	18.1%	34.2%
mid (50–85)	4,563	23.9%	37.2%
high (85–100)	1,698	48.2%	52.1%

Opt-out audit evidence

data/optout_audit_evidence.jsonl — one row per domain, 10,000 total, the union of all three splits. This file is global rather than per-split, since reservation status is independent of split assignment. It records the check output directly, not only the resulting verdict:

Field	Notes
`domain` · `checked_at` · `verdict`	`verdict` is `NO_RESERVATION_FOUND` on every row in this file. A domain showing a signal, or one that could not be confirmed clear, was excluded from the dataset (see Sampling methodology) rather than retained with a negative verdict.
`signals`	empty on every row in this file by construction; populated only for excluded domains, which are not included here
`robots_status` · `tdmrep_status` · `aitxt_status` · `llmstxt_status` · `home_status`	HTTP status per channel, or exception type if unreachable
`robots_ai_agents_blocked` · `robots_wildcard_disallow_root`	parsed from `robots.txt`
`robots_txt` · `tdmrep_json` · `ai_txt` · `llms_txt` (+ `_sha256` for each)	verbatim file content, where present and plausible (see below)

Withheld content. homepage.html is not published: it is a complete webpage rather than a machine-readable policy file, and adds no verification value beyond the fields above. Of the files that returned content, 20.6% were withheld as implausible policy-file content — some hosts return a full application shell with HTTP 200 at these paths instead of 404, producing multi-megabyte responses that are not policy files (largest observed: 5.4MB). _sha256 is recorded for every withheld file, permitting independent verification by re-fetch and hash comparison.

Procedure: optout_audit.py (live check, deterministic pattern matching, no model-based judgment) → optout_recheck.py (retry and classification of unreachable domains) → build_optout_evidence.py (this file).

Second-pass prose review, and every domain removed

The check above is deterministic pattern matching: it recognizes a reservation only via a literal token — a named AI-agent user-agent disallowed, a tdm-reservation: 1 field, or a fixed phrase matched against ai.txt/llms.txt. It does not parse robots.txt comments for meaning; comments are stripped before the file is parsed, by design. A reservation stated in prose there, without a corresponding technical rule, would not be caught by this check alone.

A second pass addressed this gap, run after initial publication: robots.txt comments (the free text after #, extracted separately from the technical rules already covered by the check above — those are not re-shown to the model, since testing found doing so caused occasional misreadings of what the parser already handles correctly) plus the full text of ai.txt/llms.txt, sent to Qwen3.5-2B with two direct questions — does the site's own text deny scraping its images for a dataset, and does it deny training AI models on its content. Either answer being non-permissive excludes the domain.

This found 281 further domains, 226 of which were part of the previously published version of this dataset. Those 226 were removed and replaced with same-stratum substitutes drawn from the same backfill pool used for the original opt-out exclusions, preserving the popularity-stratified sampling shape described above.

data/optout_rejected_evidence.jsonl — every domain excluded by either stage, 2,772 rows total: rejected_by is pattern_matcher (verdict and matched signals, as in optout_audit_evidence.jsonl) or llm_review (the model's quoted evidence and reasoning), plus ever_shipped_before_rejection marking the 226 caught only after they had already been published once. This is what makes the audit's rejection rate checkable directly rather than only claimed: a domain is either in the shipped 10,000 with a clean verdict, or in this file with a stated reason it is not.

Procedure for this pass: optout_llm_judge.py (model review) → optout_llm_merge.py (substitution) → build_optout_rejected_evidence.py (this file).

Limitations

Annotations are model-generated and not human-verified. style_tags and motif were produced by a vision-language model; sliders by an LLM. Manual inspection of a sample found shape-identification errors in approximately 20–30% of images (e.g., a dolphin identified as a bird; chevrons identified as parallelograms). Structural and typographic tags are more reliable than object-identity tags. motif and object-identity tags should be treated as weak labels.
82 blank renders (0.82%) were identified and replaced. This is a rendering failure, distinct from the labeling-accuracy issue above: the image contains no rendered content. Identified by exhaustive per-image pixel-variance screening of the full 10,000, not a sample, and replaced with same-stratum, opt-out-cleared alternatives drawn from the backfill pool (see fix_blank_images.py). No blanks remain at the applied threshold; a small number of genuinely minimal marks may score close to it.
palette_hex was removed prior to publication. The source color-extraction field contained fixed placeholder values in a majority of positions across nearly every row (one position held the constant #FF253A in all 10,000 rows). Given the risk of residual contamination in any partial correction, the field was excluded rather than shipped.
DINOv3 is trained on natural images, not vector graphics, and has no text-reading capability. The embeddings should be treated as a general-purpose visual descriptor, not one validated for logo-specific similarity or for text content.
Sparse-field coverage correlates with popularity stratum (see above).
Category and brand-name labels are inherited from third-party metadata and were not independently re-verified.
Geographic and sector distribution skews toward US/EU commercial domains. The stratified sampling corrects the popularity axis only.
caption, paired with its corresponding image, constitutes a usable image-to-design-description training pair. This is a deliberate inclusion, relevant to assessing downstream use.

Excluded fields

This release documents and measures existing logos. It excludes most material that would support training a logo-generation model: business descriptions, the compositional render schema, and the conditioning vocabulary used internally by LogoLabs models.

Two exceptions are included deliberately: motif (one sentence) and caption (a full design brief). caption is the same supervision signal an earlier LogoLabs drafter model was fine-tuned on; its inclusion was a specific, recorded decision rather than an omission from the general exclusion policy. brief_qwen8b and brief_qwen32b (a more detailed regeneration of the same underlying content) remain excluded.

Enforced by build_dataset.py and checked by verify_release.py, which validates output field names and lengths against a fixed policy so that a future edit cannot silently widen the excluded set.

Citation

@misc{logobrief10k,
  title  = {LogoBrief-10K: Design-Annotated Logo Renders with Balanced Popularity Sampling},
  author = {LogoLabs and Deleanu, Stefan-Lucian},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Logolabs/LogoBrief-10K}
}

Contact

Removals, corrections, and access questions: office@incorpo.ro

LogoBrief-10K · LogoLabs · 2026