Dsc2336/arc-annotation-demo-dataset

ARC Image Annotation Demo Dataset (205 images, COCO + YOLO)

Free sample of ARC's dual-model image annotation pipeline — the exact process behind the ARC annotation gig on Fiverr (seller: Scott C / dsc2336). One complete, real run: 205 Wikimedia Commons images labeled by one vision model, independently verified by a second model from a different provider, with disagreements adjudicated. QA numbers are published below, flaws included — the honesty is the pitch. The paid gig sells the verify pass and revisions, not an accuracy guarantee.

Headline numbers: 515 objects boxed across 11 classes · inter-model agreement on primary class before adjudication 72.2% · human spot-check (30 images): primary class 28/30 correct · bounding boxes are demo-grade — see the QA section of the run report below before judging.

Licensing: the annotations and this card are shared for evaluation; the 205 source photos are NOT redistributed here — sources.json links every image to its Wikimedia Commons page with its per-image CC license and attribution.

Files: labels/labels.json (full per-image records + labeler/verifier/adjudication trail) · labels/coco.json (COCO format) · labels/yolo/ (per-image YOLO txt + classes.txt) · labels/qa_stats.json · sources.json · arc-annotation-demo-dataset.zip (everything in one download) · the full run report follows below the line.

---

Demo annotation run — furniture taxonomy (205 images)

A working demonstration of the dual-model annotation pipeline sold in ../gig-annotation.md: every image is labeled by one vision model, checked by a second model from a different provider, and disagreements on the primary class are settled by an adjudication pass. Built entirely on free tiers ($0).

• Dataset: 205 photographs from Wikimedia Commons (see Sources below)
• Taxonomy (11 classes): bed, bicycle, bookshelf, chair, clock, desk, dining_table, lamp, refrigerator, sofa, television
• Outputs: 205 images labeled, 515 objects boxed
• Model agreement on primary class: 148 / 205 (72.2%) agreed outright; the other 57 went to adjudication
• Human spot-check (30 images): primary class right 28/30; bounding boxes are the weak point — details in the QA section, nothing hidden

Deliverable files (in labels/)

file	what it is
labels.json	full per-image records: caption, primary class, all objects (0–1000 relative coords), plus the labeler/verifier/adjudication trail for every image
coco.json	COCO-format dataset (absolute-pixel xywh boxes) for the bounding-box tier
yolo/*.txt + yolo/classes.txt	one YOLO label file per image (normalized center format)
qa_stats.json	counts: agreement %, adjudications, per-class image/object totals
_checkpoint*.jsonl	raw per-image pipeline records (the audit trail the outputs were built from)

Rebuild the outputs any time with python3 build_outputs.py; the pipeline itself is run_pipeline.py (resumable — it skips images already labeled).

Method — exactly what labeled your images

1. Labeler — Mistral Pixtral-12B (pixtral-12b-2409, 195 of 205 images). Names the primary class, writes a one-sentence caption, and draws boxes for every instance of the 11 classes. (The run started on Google Gemini 2.5 Flash, which labeled the first 10 images before its free quota ran out on 2026-10-20 and Pixtral took over. The per-image records in labels.json name the model that actually labeled each image.)
2. Verifier — a second model family, image in hand. It independently names its own primary class, then reviews the labeler's boxes and returns issues and, where needed, corrected boxes. Corrected boxes replace the originals when the two models agree on the primary class.
   • First 32 images: Qwen (qwen3.8-27b) via Groq — until that org's free-tier quota ran dry mid-run.
   • Remaining 173 images: Meta Llama 3.2 11B Vision via Cloudflare Workers AI (free neuron allocation; ≈8 neurons per verify call).
3. Adjudication. When labeler and verifier name different primary classes, a third look at the image (Pixtral for 55, Gemini for 2 before its quota died — each record names its adjudicator; both opinions are quoted in the prompt) picks the final class and final object set. 57 of 205 images (27.8%) needed it — this dataset is full of cluttered rooms and odd photos where "the main subject" is a genuine judgment call.

Why the verify pass matters, from this run's first image: the labeler boxed a lamp at roughly double its real size; the verifier flagged it and supplied a corrected box, which is what shipped. Across the run the verifier raised object-level issues on 60 of 205 images.

QA — pipeline stats

From labels/qa_stats.json:

• Primary-class agreement before adjudication: 72.2%
• Adjudications: 57
• Images where the verifier flagged ≥1 object problem: 60
• Objects per class: chair 220, lamp 54, desk 53, bookshelf 33, bed 31, clock 31, bicycle 24, dining_table 22, sofa 21, television 16, refrigerator 10 (chairs dominate because dining/classroom scenes hold many)

QA — human spot-check (30 of 205, read by a human reviewer)

Sample: random, seed 20261020 (random.sample over the finished labels). Method: the final boxes were drawn onto copies of the photos and each overlay was read against the photo. Judged on: (a) is the primary class the real main subject, (b) are the boxes tight, complete, and on real objects.

Results:

• Primary class: 28 / 30 correct (93%). The 2 errors: an ink illustration of a man in an ornate chair called "none" instead of chair; a wall-mounted dining table called desk instead of dining_table. One more (a library reading room called bookshelf where the hall clock is arguably the subject) was counted correct but is a genuine judgment call.
• Boxes fully clean: 6 / 30. Single clear subjects (a mosaic lamp, three clocks, a vintage TV, a chair close-up) plus two correctly-empty results — a refrigerator wiring diagram and an industrial refrigeration unit both correctly got primary "none" and no boxes.
• Boxes usable with minor flaws: 9 / 30 — slightly loose or slightly shifted boxes, or one small missed fixture in an otherwise good set.
• Boxes materially flawed: 15 / 30. The failure modes, in order:
  1. Loose boxes — boxes visibly larger than the object (the most common flaw; e.g. an Adirondack-chair arm, a styled bed shot).
  2. Missed instances in busy scenes — a second bed unboxed; a television in a TV-studio shot unboxed; reading lamps in a library hall unboxed.
  3. Misaligned/shifted box sets — two photos (a 1916 library document photo, a library viewed through a doorway) where several boxes sit off their objects, as if the coordinate grid drifted.
  4. Occasional phantom boxes — two small "lamp" boxes on blank wall in an attic bedroom; a stray "dining_table" label on a border-gate photo whose actual bicycle was missed.
  5. Coarse repeated boxes in dense scenes — a classroom of chair-desks and two banquet scenes got overlapping, template-like chair boxes, and in the banquet shots the dining table itself never got its own box.

Honest bottom line: the pipeline's classification is strong (93% on the spot-check) and the verify pass demonstrably catches gross errors, but box tightness and completeness in cluttered scenes is demo-grade, not production-grade — roughly half the sampled images had at least one material box problem. That is exactly why the gig listing promises a verify pass and revisions instead of an accuracy guarantee, and why throughput is capped: quality time per image is the product. A human cleanup pass on flagged images (the 60 the verifier complained about) is the obvious next improvement and is how a paid "verified" tier would be staffed.

Sources & licenses

All 205 images come from Wikimedia Commons and are freely licensed. sources.json (in this folder) maps every file to its Commons page and license. License mix across the set:

license	images
CC BY-SA 4.0	78
CC BY-SA 3.0 (incl. de/pl/cz regional ports)	32
CC BY 2.0	21
CC0	19
Public domain	19
CC BY 4.0	13
CC BY-SA 2.0	13
CC BY 3.0 (incl. de port)	8
CC BY-SA 2.5	1
CC BY 2.5	1

Attribution for any reuse lives in sources.json (per-image). The photos were fetched by Commons category search, which is noisy — several "refrigerator" results are diagrams and machinery (correctly labeled "none"), and fetch category was never treated as ground truth; every label comes from the models looking at the pixels.

Reproduce / extend

• Fetch more images: python3 fetch_images.py (Commons API, keyless)
• Run/resume labeling: python3 run_pipeline.py [--shard i --of n] (needs the environment's free-tier vision connectors; checkpoints make it crash-safe)
• Rebuild outputs + stats: python3 build_outputs.py