Nonconformance root-cause tasks from plant quality records, 16 firms

Summary

4,800 graded tasks built from real nonconformance reports and their closed CAPAs at metal fabrication, plastics and food packaging plants. Each task gives the model the defect report, inspection data and the first week of email, then asks for the root cause, the containment and the corrective action. The answer key is the CAPA the plant closed and verified, with a rubric that names the tempting wrong answers. Plants, people, customers and part numbers are replaced with placeholders.

Collected to order. We collect it from US companies for your order. Scope, volume and timing are confirmed in writing before you sign.

Pricing on request

Request scope and pricing

At a glance

Planned scope. Volume, sources and timing are confirmed in writing before you sign.

Records: 4,800graded tasks
Size: 140 MB
Tokens, estimated: 30M
Years: 2015 to 2026
Firms: 16

Owner profile: 16 US plants in metal fabrication, plastics and food packaging, 25 to 400 staff each
Industry: Manufacturing
Data type: Evaluation and task packs
Headcount: 25 to 400 staff per firm
Region: United States
Languages: English
Formats: JSONL
Updated: 2026-10-06

Sources

System	Kind	Records
Quality management systems	Databases	61K NCRs, 7.2K CAPAs
Microsoft 365 mail	Email	380K messages linked to NCRs
Shared drives	Shared files	19K inspection reports

Good for

evaluating root-cause reasoning on real manufacturing defects with a verified answer
benchmarking agents on containment versus corrective action planning
reward modeling for 8D and CAPA write-ups
testing whether a model ignores plausible distractors in inspection and email evidence

Fields

Field	Type	Description
task_id	string	Task id
plant_type	enum	metal fabrication, plastics or food packaging
difficulty	enum	standard or hard
prompt	string	The question put to the model
evidence	array<string>	NCR, inspection, log and email excerpts given to the model
reference_root_cause	string	Root cause from the verified CAPA
reference_actions	array<string>	Containment, correction and verification actually taken
rubric	array<string>	Grading points, including distractors to reject
source_ncr	string	De-identified source NCR id

Example record

{
  "task_id": "RC-02218",
  "plant_type": "metal fabrication",
  "difficulty": "hard",
  "prompt": "A customer reports weld cracks on 3 of 400 frame assemblies. Using the evidence, give the most likely root cause, the containment and the corrective action.",
  "evidence": [
    "NCR: cracks at the toe of the bracket to tube fillet weld, found at customer after powder coat",
    "Weld log: cell 4 robot, new lot of ER70S-6 wire loaded on 2nd shift 05/09",
    "Program log: WLD-118 edited 05/09 to cut cycle time, travel speed raised from 18 to 26 ipm",
    "Inspection: undercut at the toe on 5 of 12 retained parts welded 05/09 to 05/16"
  ],
  "reference_root_cause": "The program edit raised travel speed, leaving undercut and poor fusion at the weld toe, which cracked under the customer's fatigue loading.",
  "reference_actions": [
    "Contain: hold and dye penetrant test all frames welded 05/09 to 05/16",
    "Correct: restore 18 ipm and require engineering sign-off for weld program edits",
    "Verify: macro etch one part per shift for 30 days"
  ],
  "rubric": [
    "Names the program edit as the cause, not the wire lot",
    "Bounds containment by the dates the edited program ran"
  ],
  "source_ncr": "NCR-[SITE_5]-2024-0311"
}

One record, de-identified. Placeholders such as [PERSON_12] and [COMPANY_4] stay the same across the whole dataset, so threads still read correctly.

Licenses available

Exclusive
One buyer. Nobody else can license it during the exclusive term, usually 2 years.
Non-exclusive
Licensed to more than one buyer.

What ships with it

Datasheet: what it is, what it is good for, verified size, sources, years, format and limits.
Provenance file: the owner shown by code, staff notice and consent status, exclusions, mapped to the EU AI Act Article 53 training-data summary and California AB 2013.
Manifest with a hash for every file.
QC report and de-identification report.
Your own copy, with unique canary records.

How it is de-identified

Default exclusions removed, files flattened to text, identifiers found in three layers and replaced with consistent placeholders, and the result measured against a 99% bar for each identifier type in each source.

The full method