SAgoge: Multimodal SVG Dataset

• SAgoge is a comprehensive multimodal vector-graphic dataset defined by deep hierarchical structure and broad SVG modality coverage including icons, illustrations, diagrams, and animations.
• It employs advanced annotation schemas powered by state-of-the-art language models to support tasks in SVG understanding, editing, and generation with stratified difficulties.
• The dataset scales to 16M samples across diverse domains, providing a unified benchmark (SArena) that drives research in long-sequence code modeling and cross-modal translation.

SAgoge is a large-scale, comprehensive multimodal dataset for vector-graphic research, specifically targeting Scalable Vector Graphics (SVG) understanding, editing, and generation. Developed as the foundational resource for the InternSVG model suite and SArena benchmark, SAgoge introduces unprecedented scale, modality coverage, and structural richness for SVG tasks, unifying domains such as static icons, long-sequence illustrations, scientific diagrams, and dynamic animations (Wang et al., 13 Oct 2025).

1. Scope, Scale, and Structural Properties

SAgoge contains approximately 16 million training samples drawn from four distinct vector-graphic modalities:

Modality	Raw SVG Files	Derived Task Samples	Avg. SVG Length (tokens)
Static Icons	2.8 M	11 M	846
Long-sequence Illustrations	600 K	1.6 M	8,673
Scientific Diagrams	1.7 M	3.4 M	1,752
Dynamic Animations	61 K	122 K	847

The dataset preserves deep hierarchical code structure—retaining SVG <g> (group) tags and nested primitives—enabling scenes comprised of hundreds or thousands of graphical elements. This structural fidelity exceeds prior corpora, which are typically limited to flat icon-level representations.

2. Annotation Schema and Attribute Inventory

Annotations were automatically produced using high-capacity multimodal LLMs (GPT-4o, Qwen2.5-VL, InternVL3 for image supervisions; Gemini-2.0 Flash for video). SAgoge distinguishes three primary families of tasks, each supported by systematic annotation schemes:

• SVG Understanding:
  • Description: Free-form semantic and geometric text.
  • Multiple-Choice QA: Four questions per sample covering color, geometry, object count, and semantic identity, with options (A–D).
• Editing Tasks (ten subtasks):
  • Syntactic (color swap, add stroke, translation, scale, rotation, flip, opacity, crop)
  • Semantic (semantic color editing, style transfer)
• Generation Tasks:
  • Input/Output: Text/image input for static SVG, text/video input for animation, output as SVG or “SANI” (SVG Animation).

A tokenizer with 55 tag tokens, 42 attribute tokens, 247 integer tokens (–128 to 128), and 100 fractional tokens (.00–.99) encodes SVG code. This reduces average sequence length by 20–30% while retaining fine-grained numeric precision.

3. Data Acquisition, Normalization, and Quality Control

SAgoge aggregates data from diverse public and synthetic origins:

• Icons and some illustrations/animations: Public repositories (Iconfont, SVGRepo, OpenClipart)
• Scientific diagrams: PubChem SDF to SVG via Open Babel
• Illustrations: A synthetic chain (GPT-4o prompts → FLUX vector image generator with vector LoRA → VTracer vectorizer)
• Animations: Claude-Sonnet-4 code synthesis to SMIL-compliant SVG animations

Normalization includes uniform resizing of viewBoxes to 128×128 for stable coordinate ranges, code simplification (removal of metadata, redundant elements), and preservation of hierarchical tags and core primitives. Manual screening removes corrupted/ambiguous files for the SArena evaluation splits, while non-rendering or annotation-failures in training are dropped or corrected.

Sampling and curriculum balancing proceed in two stages: initially restricting to Icon and Chemistry domains (simple), then expanding to Illustration and Animation domains and adjusting domain proportion.

4. Task Design, Evaluation Metrics, and Difficulty Stratification

SAgoge supports a broad spectrum of tasks stratified by domain and complexity:

• Understanding:
  • Easy tasks (color, count, geometry) evaluated by accuracy.
  • Hard tasks (semantic identity) require inference from SVG code structure.
• Editing:
  • Simple subtasks (e.g., color swap, transform) versus hard (semantic color swap, style transfer).
  • Task grading employs metrics such as PSNR and LPIPS to distinguish between syntactic and semantic edit complexity.
• Generation:
  • Includes domains such as icons (short sequences), illustrations (long-sequence), chemistry (structure-specific), and animation (temporal sequences).
  • Input modalities: text, image (for static graphics), text and video (for animations).
  • Output modalities: SVG for static, “SANI” for animations.

No per-sample numeric difficulty score exists, but the SArena benchmark stratifies editing tasks as “simple” or “hard.”

Evaluation metrics are standardized by modality and include: accuracy (%), DINO, SSIM, LPIPS, PSNR, FID, FID-CLIP, CLIP-T2I or CLIP-T2V, CLIP-I2I or CLIP-V2V, and FVD for animation. Black outputs are substituted for unrenderable SVGs during metric computation.

5. Data Splitting Strategy and Benchmarking

SAgoge is solely intended for large-scale pretraining. For evaluation, the SArena benchmark provides held-out validation and test sets across domain/task axes:

Domain	QA	Editing	Text-to-SVG	Image-to-SVG	Text-to-SANI	Video-to-SANI
Icon	6,012	2,000	6,013	6,013	—	—
Illustration	—	—	2,001	2,001	—	—
Chemistry	—	—	3,003	3,003	—	—
Animation	—	—	—	—	504	504

This structure enables comprehensive, domain-aligned, and difficulty-stratified assessment of model capabilities. SAgoge itself has no official public split for train/validation/test; all benchmarking leverages SArena’s curated sets.

6. Comparison with Prior SVG and Vector Datasets

Relative to historical SVG datasets, SAgoge sets new benchmarks in scale, scope, and structural granularity:

Dataset	Samples (M)	Modalities	Hierarchy/Attributes
ColorSVG-100K	0.1	1 (icon)	Flat/limited
DeepSVG	0.1	1 (icon)	Flat/limited
UniSVG	0.525	1–2	Basic groupings
SVGX	1	1–2	Basic groupings
SVG-Stack	2.2	2	Limited hierarchy
MMSVG	2	2	Limited hierarchy
SAgoge	16	4	Deep hierarchical, rich attr

Formally, SAgoge achieves $$(\mathcal{D},\,\mathcal{M})_{\rm SAgoge} = (16\mathrm{M},4)$$, eclipsing all predecessors in both dimensions.

SAgoge is the first dataset to unify:

• Understanding, editing, and generation tasks;
• Four vector-graphic modalities within a single corpus;
• Detailed attribute and group hierarchies encoding scene complexity.

7. Practical Applications and Known Limitations

Applications

• Training LLMs and multimodal LLMs for end-to-end SVG reasoning across comprehension, manipulation, and synthesis.
• Research in long-sequence code modeling (up to 8,000+ tokens/sample).
• Cross-modal translation tasks, including image/video-to-vector and text-to-animation generation.
• Development in domain-specific graphics, notably scientific diagrams and chemical structures.

Limitations

• Synthetic illustrations and animations may diverge from real-world data distributions.
• LLM-generated annotations introduce label noise; annotation quality is contingent on the underlying annotators’ capacity.
• Code simplification strips metadata fields (IDs, comments), which restricts utility for use-cases dependent on provenance or lineage information.
• SAgoge provides no official validation/test splits, requiring reliance on the companion SArena benchmark for evaluation.
• Training on the full dataset incurs substantial computational overhead.

---

SAgoge represents the largest, most structurally expressive and task-diverse SVG corpus to date, constituting a critical path toward unified, multimodal SVG modeling (Wang et al., 13 Oct 2025).