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.
187-annotated
Chain-of-thought defect inspection on metal part — 846 items (564 good + 282 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 a teacher reasoning chain landing on the gold 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 scratches, anomalous, parts mismatch, total rust, bend and parts mismatch, major rust, defective painting, hole and region is a coarse position phrase (3x3 grid or a span phrase),
both derived deterministically from the human segmentation mask of AI4Manufacturing/187.
Coverage: anomalous 69, bend and parts mismatch 17, defective painting 13, hole 12, major rust 14, parts mismatch 48, scratches 86, total rust 23.
Reasoning channel
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.
Grounded traces
The reasoning_grounded column was removed 2026-07-21 per the corpus grounded-trace doctrine
(decided 2026-07-18): for classification-shape tasks the trace can only restate the verdict (with
unrequested coordinates) — geometry stays in metadata; traces are pure functions of the GT and the
committed generator (forge_model annotate/cot_anom/) and can be re-rendered on demand. Default SFT
view: query -> reasoning. RLVR uses query + annot.
Roles & baseline
Roles: reasoning is the SFT imitation target — it ends FINAL ANSWER: <annot> (byte-exact),
the segment a student emits; annot is the machine-parseable gold / reward key, not an output-format target;
answer-only exact-match RL needs only query + annot (no reasoning). Items are ~2:1
good:defective, so a blind "always good" scores 67% on the label alone — the defect type +
coarse location is what the task actually grades.
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 846 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:
187-grounding,187-region,187-mcq.
| field | meaning |
|---|---|
query |
inspection prompt (diverse paraphrase pool; independent of the gold) |
image |
the metal part photo (native 1024x1024) |
annot |
gold JSON: defects (type + region) + label |
reasoning |
teacher CoT (gpt-5.4-mini) + 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/187
(binarized at gray>127 (clean 0/255 masks); the defect TYPE comes from metadata (one per image, raw underscores normalized to spaces, raw kept in metadata.defect_type_raw)). 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 187).
- Downloads last month
- -