You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.

Log in or Sign Up to review the conditions and access this dataset content.

184-mcq

Mask-grounded multiple choice (Set-of-Mark style) for commutator surface defect localization — 50 items, derived deterministically from the binary segmentation masks of AI4Manufacturing/184. Exact-match gradable → SFT and RLVR-ready.

Task

One item per eligible defective record. image is a 2×2 grid of views A–D of the same photo, each overlaying one red candidate region mask. Exactly one view overlays the defect accurately — right location and extent. annot is the correct letter. The three negatives per item are hard by construction (shift / fliplr / flipud / rot180 / dilate ≥2.5× / erode), each guaranteed wrong (IoU vs truth < 0.35 except dilate), panels mutually distinct (pairwise IoU < 0.7). Gold letters: A 13 / B 17 / C 8 / D 12. Coverage: defect 50.

Records are skipped (confidence over coverage) when the gold mask is under ~30 visible px after panel downscale, fewer than 3 sound negatives are constructible, or the mask covers >35% of the frame. Skipped defects remain fully covered by 184-grounding / 184-region.

field meaning
query names the defect (binary); "answer with the letter only"
image 2×2 composite, panels A–D
annot A / B / C / D
reasoning null — deterministic
cate / task B / T-B2
metadata source, image_sha256, defect_type, gold_letter, panel_tags, area_pct

Provenance

Built deterministically (no LLM/teacher; reasoning is null) from AI4Manufacturing/184 — KolektorSDD (v1) electrical-commutator surface inspection — 383 photos (deduped from 399 by sha256), binary good vs anomalous (thin surface cracks); each anomalous image has a paired binary pixel segmentation mask (binarized at gray>40 (source masks ~0/255)). Generator: annotate/184/build_184_derived.py in forge_model; machine gates annotate/184/verify_184.py (all green at build: boxes/gold re-derived byte-identical, clean regions contain zero defect pixels, mcq gold = the true panel, no leaked-artifact vocabulary, image set sha256-disjoint from all other cached AI4Manufacturing datasets).

Resolution. Images are native (~500×1260; not downscaled); all coordinates live in that pixel space (see metadata.image_wh).

Query diversity. query is drawn from a fixed pool of surface variants (paraphrases preserving the task + answer format), selected by an independent per-record hash; a machine gate confirms no template correlates with the gold.

The repository name is an internal task code (the source dataset's code is 184).

Downloads last month
2