aixonlab/svg-sft-v1

SVG SFT v1

Chat-format supervised fine-tuning data for teaching a small LLM (target: Qwen/Qwen3.5-4B) to write SVG code from a natural-language prompt. Every assistant turn contains only an <svg>…</svg> document.

• 78,620 examples: 77,865 train / 755 validation (about 1%, stratified by category, deduplicated against train).
• Static icons, emoji, flat/line/gradient illustrations, complex scenes and animated SVG (SMIL and CSS).
• The SVG is at most 6,000 Qwen3.5 tokens (and at most 16k characters).

Format

{"messages": [{"role": "user", "content": "<prompt>"}, {"role": "assistant", "content": "<svg ...>...</svg>"}],
 "source": "iconify", "license": "MIT", "category": "icon",
 "id": "iconify:tabler:bell", "author": "...", "source_url": "...", "animated": false, "n_tokens": 212}

category is one of icon, flat_illustration, line_art_bw, shaded_gradient, emoji, animated, complex_scene.

Composition

By category

category	rows
icon	30,000
flat_illustration	12,000
animated	11,000
line_art_bw	10,000
shaded_gradient	8,911
emoji	6,000
complex_scene	709

By source

source	rows
iconify	20,352
sagoge	18,750
svgrepo	16,399
clker	10,761
openclipart	6,529
svg-sophia	5,829

By license (from each row's license field)

license	rows
Apache-2.0	30,833
CC0-1.0	21,464
MIT	17,023
CC-BY-4.0	4,428
Public-Domain	4,372
ISC	323
Unlicense	110
BSD-3-Clause	67

Category × source

category	iconify	sagoge	svgrepo	clker	openclipart	svg-sophia	total
icon	10,909	4,091	9,545	0	0	5,455	30,000
flat_illustration	0	3,004	2,251	4,164	2,581	0	12,000
animated	1,627	9,363	0	2	8	0	11,000
line_art_bw	0	794	4,380	2,661	2,165	0	10,000
shaded_gradient	1,816	1,301	177	3,638	1,605	374	8,911
emoji	6,000	0	0	0	0	0	6,000
complex_scene	0	197	46	296	170	0	709

Animated rows: 10,997 use SMIL and 1 use CSS @keyframes. 139 Iconify sets and 1274 SVG Repo packs are represented.

SVG token length (Qwen3.5-4B tokenizer, assistant content only)

category	n	mean	p50	p95	max
icon	30,000	1087.5	559	3674	6000
flat_illustration	12,000	2533.0	2363	5322	6000
line_art_bw	10,000	1880.2	1394	4950	5993
shaded_gradient	8,911	2421.2	2149	5333	6000
emoji	6,000	2170.8	1853	5101	5998
animated	11,000	801.9	657	2077	5966
complex_scene	709	4808.7	5105	5918	5999
all	78,620	1636.4	1013	4842	6000

Prompt tokens: mean 30.1, p50 15, p95 109, max 203.

Sources and licenses

source	upstream	license handling
iconify	iconify/icon-sets JSON	Only sets whose SPDX license in the set's own info.license is MIT, Apache-2.0, CC0-1.0, ISC, Unlicense or BSD-3-Clause. Sets in the Logos, Flags/Maps and Programming categories are excluded, and so are brand-named icons. This includes the animated sets line-md, meteocons, svg-spinners and eos-icons (MIT). At most 330 icons are taken per set (1,800 per emoji set) to keep styles diverse. author and source_url give the set author.
svgrepo	nyuuzyou/svgrepo	Only rows whose in-row license is MIT, PD, CC0, Apache or CC Attribution. The Logo, CC-NC, GPL, MPL and other splits are dropped. Packs matching logo, brand, emoji (twemoji, emojione, openmoji), flag, social, payment, team, vscode, programming, etc. are excluded, and so are brand-named titles and tags. At most 120 rows are taken per pack. author = license_owner.
clker	nyuuzyou/clker-svg	CC0 per the dataset card. Rows carry no license field. This is a sample from 8 of the 26 shards.
openclipart	nyuuzyou/openclipart	The in-row license must be CC0. This is a sample from 3 of the 18 shards (≤40k raw chars). author = artist_name.
sagoge	InternSVG/SAgoge (Apache-2.0)	Captioned stage2/animation/text2sani (animated, SMIL), plus spread byte-range samples of stage2/icon/generation/text2svg and stage2/illustration/text2svg. Caption rows mentioning logo, brand, trademark or wordmark are dropped.
svg-sophia	InternSVG/SVG-Sophia CoT-SFT text2svg (Apache-2.0)	Only the final svg block is kept. The <think> CoT and the "use this format" prompt preamble are removed.

Deliberately excluded: MMSVG / OmniSVG, SVGX, LottieGPT (non-commercial); starvector svg-stack / text2svg-stack (license undeclared); SAgoge video2sani (its prompts refer to a video input); svgrepo logo, cc_nc, GPL, MPL and emoji packs; Iconify CC-BY, CC-BY-SA, GPL and OFL sets.

Attribution (CC-BY rows)

Rows with license = CC-BY-4.0 come from SVG Repo's "CC Attribution" split. Their author field holds the SVG Repo license_owner and source_url links the original page. The file attribution_cc_by.tsv lists every one. If you redistribute these SVGs or derivatives, credit the authors as CC-BY requires. Iconify rows keep the set author and URL. MIT, Apache and ISC sets require their license notice to be kept when the icons themselves are redistributed. See each set's repository (source_url).

Cleaning pipeline

1. Safety and parsing. Parsed with lxml (no entity resolution, no network, no DTD). Rows are rejected if they have DOCTYPE entities, <script>, <image> or data: URIs (embedded rasters), fail to parse, or are too long.
2. Removing cruft. Comments, processing instructions, <metadata>, <title>, <desc> and <foreignObject> are removed, along with every non-SVG-namespace element and attribute (Inkscape, Sodipodi, RDF, Adobe, etc.; xlink:href is kept). Also removed: data-*, version, enable-background, inkscape-only style properties, unreferenced ids (kept on animated SVGs), unused unnamed gradients/clipPaths/masks/patterns, empty <defs>, and whitespace-only text nodes.
3. viewBox. Always present. If it is missing it is derived from width and height, with units converted to px at 96 dpi. Root width, height, x and y are removed so outputs are resolution-independent.
4. Number normalization. Decimals are rounded to 2 places (3–4 places for viewBoxes under 10 or 1 units). Leading zeros are dropped. Compact arc-flag tokens are left untouched.
5. Render validation (static). Rendered with resvg on a transparent background. An SVG is rejected if alpha coverage is below 0.4% (blank) or if coverage is above 97% with near-zero color variance (a solid block). The cleaned render is compared with the render of the original, and the SVG is rejected if the mean absolute RGBA difference is above 6/255. Animated SVGs only need to parse (resvg cannot play SMIL/CSS).
6. Deduplication. In priority order (Iconify > SVG Repo > Openclipart > Clker > SVG-Sophia > SAgoge, permissive before CC-BY), using three checks:
   • an exact hash of the whitespace-stripped normalized SVG
   • a 256-bit dHash of the 96 px render (static)
   • for animated SVGs, a scale-invariant hash with all numbers divided by the viewBox width, which catches SAgoge's 128-px rescaled copies of line-md
7. Token filter. At most 6,000 Qwen3.5 tokens.
8. Category balancing. Weighted round-robin across sources within each category, using fixed per-category targets.
9. Split. 1% stratified validation. The script asserts that train and val share no exact hash, render hash, animation hash or prompt.

Drop reasons per source are recorded in stats.json (clean_rejects_by_source, assemble_drops_by_source).

Prompts

• Captioned sources (SAgoge, SVG-Sophia) reuse the upstream caption. The system-style preamble ("You are an expert…") and any CoT format request are removed, and the caption is wrapped in varied request templates. For the animation captions, "The video shows…" phrasing and sentences that call the animation static are removed. A motion sentence derived from the SVG itself is appended, for example "its strokes draw themselves in", "an element rotates", "loops continuously" or "plays once", together with the technique (SMIL or CSS).
• Metadata sources (Iconify, SVG Repo, Clker, Openclipart) get template prompts built from the icon name, title, tags, the Iconify category or the Openclipart description, varying from 2-word queries to detailed requests. Style cues are computed from the SVG, never chosen at random. "gradient-shaded" appears only if gradients or filters exist, "black and white" only if the palette is gray-scale, "outline" only if strokes dominate, the viewBox only as it actually is, and colors only as nearest names for the colors actually used.
• No LLM was used to write prompts or SVGs in this build.

Caveats

• Provenance of InternSVG data. SAgoge and SVG-Sophia are Apache-2.0 as published, but their upstream SVG sources are not documented. The license covers the release, and the original icon authorship is unknown. SVG-Sophia's CoT text was generated with Claude-Sonnet-4.5 according to its authors. Only its SVG is used here (no CoT), but the SVGs and captions of that dataset passed through that pipeline. Filter source in {"sagoge","svg-sophia"} out if you need a strictly first-party-licensed subset.
• SAgoge illustrations are auto-traced (many short polygon segments). They make up most of complex_scene.
• Template prompts are less diverse than human prompts. Clker and Openclipart titles can be noisy.
• Logos and brands were filtered by keyword. Some trademarked shapes may remain. Please report them.
• Text in SVGs is not rendered during validation (no system fonts), so text-only graphics were dropped as blank.
• Animated SVGs were not frame-rendered. They are syntactically valid but not visually verified.
• This is not legal advice. Check the per-row license before commercial redistribution of the SVGs themselves.

Build

Scripts: scripts/collect.py, scripts/clean.py, scripts/svgclean.py, scripts/prompts.py and scripts/assemble.py in the svg-slm project (aixonlab), seeded and deterministic.