179-annotated / README.md
weipang142857's picture
remove reasoning_grounded (grounded-trace doctrine conformance; audit 2026-07-21; deterministic column drop, v1<->v2 diff-verified)
a0d2d67 verified
|
Raw
History Blame Contribute Delete
3.78 kB
metadata
tags:
  - smart-manufacturing
  - sft
  - chain-of-thought
  - industrial
  - vision
license: other
pretty_name: 179-annotated
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.

179-annotated

Chain-of-thought (CoT) reasoning annotations for aero-engine turbine-blade defect inspection (AeBAD) — 2,160 items (1,011 good + 1,149 defective), the reasoning-augmented sibling of 179-grounding / 179-region / 179-mcq, derived from AI4Manufacturing/179.

Task

Grade each blade good or defective; if defective, name every defect type and its coarse region. Defect classes: ablation, breakdown, fracture, groove.

Composition

2,160 rows — good 1,011 · defective 1,149 (ablation 169, breakdown 329, fracture 389, groove 262).

Schema (8 columns)

field meaning
query student question (prose; states the closed defect-class list)
image the raw inspection photo
annot machine-parseable gold JSON: `{"defects": [{"region", "type"}, ...], "label": "good"
reasoning teacher chain-of-thought ending FINAL ANSWER: ...
mask pixel GT mask (defective rows)
cate / task B / T-B2
metadata source, category, image_sha256, image_path, split, …

How the reasoning was made

  • Teacher: gpt-5.4-mini (OpenAI Batch API), English, gold-conditioned rationalization — the teacher writes forward reasoning that lands on the given human gold; it never re-solves the image and never rejection-samples the answer. The gold is authoritative; the reasoning explains it.
  • No leakage: the prompt forbids naming any grounding artifact (mask / ground-truth / reference / coordinates / "the given answer"); a regex gate confirms 0 leaked-vocabulary hits.

Roles

Roles: reasoning is the SFT imitation target — its FINAL ANSWER segment is the model-facing answer format; annot is the machine-parseable gold used for verification and reward parsing, not an output-format target.

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/mask; traces are pure functions of the GT and the committed generator (forge_model annotate/) and can be re-rendered on demand. Default SFT view: query -> reasoning. RLVR uses query + annot.

Faithfulness gate (Stage 7)

  • 0 leaked-vocabulary hits across the set.
  • 0 hallucinated defects on good images — the key failure mode (a clean part narrated as defective).
  • Gold-conditioned faithfulness judge (gpt-5.6-terra, a different model family from the teacher) over a 304-row good-heavy calibration sample: 1.7% flag rate, 0% on good rows. Every flagged case was either a judge conflation on a correct row or was fixed and re-verified (see the source-specific notes).

Provenance / reproduction

Built from the source dataset's human annotations by forge_model annotate/ (builders + shared annotate/teacher_prompt.py + annotate/cot_batch/ pipeline). Public, manual access review (gated=manual).