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.
188-annotated
Chain-of-thought defect inspection on mobile phone screen — 1219 items (20 good + 1199 defective). The model reads the photo and must judge good vs defective and, if defective, name each defect type and coarse location. Every item carries two reasoning channels landing on the same answer.
Task
query asks for a good/defective verdict plus, if defective, the defect type(s) and location(s).
annot is the gold answer as JSON — {"defects": [{"region": ..., "type": ...}], "label": "good"|"defective"}
— where type is one of oil, scratch, stain and region is a coarse position phrase (3x3 grid or a span phrase),
both derived deterministically from the human segmentation mask of AI4Manufacturing/188.
Coverage: oil 400, scratch 399, stain 400.
Two reasoning channels
reasoning— natural-language inspection written by a teacher LLM (gpt-5.4-mini), gold-conditioned (it rationalizes the human-verified answer, forward from the visible image), endingFINAL ANSWER: <gold>. It never references a mask/annotation/reference — it reads as an independent visual inspection.reasoning_grounded— a deterministic, code-templated chain that cites the exact defect boxes[x,y,w,h]and centers from the mask geometry, ending at the sameFINAL ANSWER. Faithful by construction; serves coordinate-native students and exact-match RL.
Quality
Golds are a pure function of the human masks (re-derived byte-identical at assembly; gold-identity asserted on every row). Defects were confirmed visible at teacher feed-resolution before generation. Teacher output was scanned for answer/artifact leakage (0 leaks over all 1219 rows) and truncation (0). A gold-conditioned faithfulness audit (gpt-5.6-terra, an independent model family) sampled the corpus good-heavy (hallucinated-defect-on-good is the main risk); see the model card discussion.
Companion tasks for the same source:
188-grounding,188-region,188-mcq.
| field | meaning |
|---|---|
query |
inspection prompt (diverse paraphrase pool; independent of the gold) |
image |
the mobile phone screen photo (native 1920x1080) |
annot |
gold JSON: defects (type + region) + label |
reasoning |
teacher CoT (gpt-5.4-mini) + FINAL ANSWER |
reasoning_grounded |
deterministic box-cited CoT + FINAL ANSWER |
cate / task |
B / T-B2 |
metadata |
source, image_sha256, image_wh, defect_type, n_instances, regions, ... |
Provenance
Golds built deterministically from the binary masks of AI4Manufacturing/188
(binarized from the palette mask at index>0 (defect TYPE lives in the palette index; LOCATION in the mask)). Reasoning: gpt-5.4-mini (gold-conditioned rationalization, not re-solving) via the
OpenAI Batch API; deterministic channel + assembly by annotate/cot_anom/ in forge_model.
Faithfulness audit: gpt-5.6-terra (independent family; the EPHONE gate key is never used for generation).
Split not pre-cut (choose train/test freely). Repository name is an internal task code (source code 188).
- Downloads last month
- -