188 / README.md
Y-xvan's picture
188: MSD -> T-B1 (unified SFT; viewer-friendly row groups)
466c30c verified
|
Raw
History Blame Contribute Delete
3.58 kB
metadata
tags:
  - smart-manufacturing
  - sft
  - industrial
  - vision
license: other
pretty_name: '188'
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
extra_gated_prompt: >-
  This dataset is released for **research use**. Access is reviewed and granted
  **manually** by the maintainers. Please state your name, affiliation, and
  intended use.

188

Mobile-phone screen surface anomaly detection (3 defect types; segmentation GT). Category B, task T-B1, in the unified Smart-Manufacturing SFT schema.

The repository name is an internal task code. See Provenance below for the underlying dataset.

Records

1,220 records (test=1200 · train=20). Pixel masks are embedded as a mask image column.

Unified SFT schema

field type meaning
query str the question / instruction (model input)
image Image the input image (bytes embedded)
annot str the answer — for this dataset: plain-text {label, defect_type}{good, null} or {anomalous, <defect>} (one of oil/scratch/stain). The palette-mode segmentation mask is deferred localization GT, with seg info (mask_path, defect_area_fraction) in metadata — see Task, mask & split below
reasoning null no native CoT in these datasets
cate "B" SFT category
task "T-xx" unified task id
metadata str (JSON) split, provenance, image_path, image_sha256 (dedup key)
mask Image | null (T-B1/T-B2 only) the pixel ground-truth mask, bytes embedded
masks list[Image] (D21 only) multi-region masks

Task, mask & split

What this is. MSD (Mobile-phone Screen surface Defect; jianzhang96/MSD) — MVTec-style unsupervised anomaly detection & segmentation of phone-screen surfaces (1920×1080, industrial camera). This repo uses the MSD-US package: 20 defect-free training images + 1,200 defective test images across 3 defect types (oil, scratch, stain; 400 each), each with a pixel-level segmentation mask. (The README describes a "PASCAL VOC" packaging; the MSD-US package we use ships no VOC XML — it is AD/segmentation with masks.)

Query & answer (this repo's SFT task). query is our own instruction template (the dataset ships no question); it names the closed set of 3 defect types and asks for the label + type. annot = plain-text {good, null} or {anomalous, <defect>} (one of oil/scratch/stain), the type taken from the source folder.

Mask (deferred localization GT). Each defective image ships a palette-mode segmentation mask (matched 1:1 by image stem under test/ground_truth/), with mask_path + defect_area_fraction in metadata; defect-free images have mask=null. A text-output model cannot emit a pixel mask, so localization is deferred.

Split. train = 20 defect-free images; test = 1,200 defective (oil/scratch/stain, 400 each). This is the unsupervised protocol — all defect-free images are used for training, so test is defect-only.

Provenance

Underlying dataset: MSD. Upstream license: GPL-3.0 (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion script: 188/convert_d88.py, published with publish/push_to_hf.py, both in AI4Manufacturing/forge_model.

Overlap / de-duplication (§8)

None notable. Each record carries metadata.image_sha256 so overlapping images can be kept entirely on one side of a train/eval split.