Manufacturing quality and inspection records for AI training

A manufacturing quality dataset is a plant's record of how parts were checked and what happened when they failed: inspection results and measurements, nonconformance reports, material review dispositions, corrective actions and 8D investigations, often with defect images. SourceX sources these records from the quality and production systems of established manufacturers, typically spanning years of operations, pseudonymizes customer, supplier and part identifiers, and leaves customer-owned drawings out unless the customer authorizes them.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Inspection results, nonconformances, dispositions and corrective actions from production plants
Typical systems: SAP QM, ETQ Reliance, MasterControl, TrackWise, Siemens Opcenter, InfinityQS
Typical history: Operations records spanning years; varies by partner
Modality: Structured QMS and measurement records, PDF reports and defect images
Delivery formats: Parquet or JSONL tables with linked images and PDFs, as agreed
Preparation: Customer, supplier, part and employee identifiers pseudonymized; drawings excluded unless authorized
Licensing: Non-exclusive, or exclusive for an agreed snapshot; field-of-use limits can apply
Availability: Depends on partner plants holding matching records; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
ncr_id	string	Pseudonymous nonconformance ID linking inspections, dispositions, corrective actions and images.
detected_at	timestamp	When the nonconformance was found, in UTC, with the shift where recorded.
detection_point	enum	Receiving inspection, in-process, final inspection, internal audit or customer return.
part	object	Pseudonymous part ID, drawing revision, operation and work center; customer tokenized.
lot	object	Lot, batch or serial range, with quantities inspected and rejected.
defect	object	The plant's defect code, category and severity, plus the inspector's free-text description.
measurements	array	Characteristic, nominal, spec limits, measured value, unit, gauge ID and result.
spc_context	object	Control chart type, control limits and the rule violation that triggered a hold.
images	array	Defect photos and microscope or machine-vision frames, with any region the inspector marked.
disposition	object	Material review decision — use as is, rework, repair, scrap or return to supplier — with approver role.
capa	object	Linked corrective action or 8D, with containment, root cause, actions and effectiveness check.
supplier	object	Pseudonymous supplier ID and any supplier corrective action request (SCAR).
documents	array	Inspection reports, first article reports, 8D write-ups and certificates of conformance as PDFs.

Example record

{
  "ncr_id": "ncr_4e17b2",
  "detected_at": "2024-05-06T07:48:00Z",
  "shift": "B",
  "detection_point": "in_process",
  "part": { "part_id": "prt_9c31", "drawing_rev": "D", "customer": "[CUSTOMER_ID]",
            "operation": "op40_finish_bore", "work_center": "cnc_07" },
  "lot": { "lot_id": "lot_7a0e", "qty_inspected": 32, "qty_rejected": 5 },
  "defect": { "code": "DIM-OOT", "category": "dimensional", "severity": "major",
              "text": "Bore A undersize on parts from spindle 2. Finish OK." },
  "measurements": [
    { "serial": "sn_0412", "char": "bore_A_dia", "nominal": 12.000, "lsl": 11.990,
      "usl": 12.010, "value": 11.986, "unit": "mm", "gauge": "g_air_018", "result": "fail" },
    { "serial": "sn_0413", "char": "bore_A_dia", "nominal": 12.000, "lsl": 11.990,
      "usl": 12.010, "value": 11.992, "unit": "mm", "gauge": "g_air_018", "result": "pass" }
  ],
  "spc_context": { "chart": "xbar_r", "rule": "trend_7_decreasing", "hold": true },
  "images": [ { "image_id": "img_0f2d", "type": "microscope", "bbox_xywh": [412, 288, 96, 80] } ],
  "disposition": { "decision": "rework", "approver_role": "quality_engineer",
                   "decided_at": "2024-05-07T14:10:00Z" },
  "capa": {
    "capa_id": "capa_21d8",
    "method": "8D",
    "containment": "100% bore check on WIP from spindle 2 until tool change",
    "root_cause": "Boring insert worn past limit; tool-life counter set above the insert maker's rating",
    "actions": ["Tool-life limit lowered", "Insert check added to setup sheet"],
    "effectiveness": { "checked_at": "2024-07-01", "result": "effective", "recurred": false }
  },
  "supplier": null,
  "documents": ["insp_rpt_5512.pdf", "8d_capa_21d8.pdf"],
  "closed_at": "2024-07-03T16:20:00Z"
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train investigation and CAPA agents

Linked nonconformance, disposition and 8D records show how engineers moved from a symptom to a contained, root-caused and verified fix, the path an investigation agent must follow.

Build visual inspection models with production labels

Defect images tied to the plant's defect codes and dispositions come from real lighting, fixtures and part variation, including borderline parts that inspectors disagreed about.

Detect process drift from measurement histories

Subgrouped measurements with spec limits, gauge IDs and hold events let you train and test models that flag drift before parts go out of tolerance.

Evaluate root-cause reasoning

Closed cases with known root causes and effectiveness results become held-out test items: given the evidence available at the time, does the model reach the conclusion that held up?

Extract data from quality documents

Inspection reports, first article reports and certificates of conformance are semi-structured PDFs with tables, stamps and signatures, a realistic document-AI test.

Use-case guides: Robotics and embodied AI, Enterprise and computer-use agents

What makes this data valuable

Closed-loop linkage

Nonconformances linked to a disposition, root cause and verified corrective action capture the whole decision.

Spec limits with every measurement

Nominals, tolerances, units and gauge IDs let you recompute pass or fail and process capability.

Effectiveness outcomes

Recurrence checks after a corrective action closes show whether the stated root cause was right.

Stable defect taxonomy

Codes kept consistent over years, or supplied with a mapping table, make long histories comparable.

Traceability context

Lot, machine, shift and supplier fields tie each defect to the conditions that produced it.

Records the standard requires

Plants working to ISO 9001, IATF 16949 or AS9100 must document nonconformities and corrective actions, so records follow a defined process.

What a quality record holds that a defect image does not

A plant's quality record holds the decisions made after a defect turned up — containment, disposition, root cause and whether the fix held — which no defect image shows. Public manufacturing datasets tend to isolate one signal, such as defect images from a lab setup or sensor traces from a test rig. They show what a defect looks like, not who quarantined the suspect lot, why the review board chose rework over scrap, or which root-cause hypotheses were tested and dropped.

That chain is engineering judgment exercised under production pressure, and it comes with an outcome attached. An effectiveness check, or a repeat nonconformance on the same characteristic months later, tells you whether the stated root cause was real. Few other sources pair reasoning with a verifiable result.

The records have known weaknesses too. They are written for auditors and customers as much as for engineers, so some investigations stop at a convenient answer. Defects that escaped inspection surface only later, as complaints or returns, and near-misses nobody logged leave no trace. A good scope measures these properties rather than discarding the data over them.

Keeping measurements useful when drawings are confidential

When a customer owns the part design, the dataset has to keep measurements meaningful without exposing the drawing. Removing the drawing files is simple. Removing every number derived from them is not: nominal dimensions, tolerances, characteristic names and even the sequence of operations can reveal the design, yet a measurement stripped of its spec limits is close to useless for training.

There are three common ways through. The scope can be limited to the manufacturer's own products. Customers willing to take part can authorize their records. Or measurements can be rescaled before delivery so that each value is a position within its tolerance band, with 0 at the lower limit and 1 at the upper. That keeps pass or fail, drift and process capability while hiding the actual dimensions. Each option changes what a model can learn, so SourceX settles it per dataset during qualification, and the sample you review reflects the choice.

Images need the same care. Part markings, customer logos, fixture labels and screens in the background can reveal the customer even after every structured field has been pseudonymized.

What to check before licensing

Ask what share of the records concern customer-owned parts and how each customer's authorization was obtained. Check sample free text and file names for customer part numbers.
Ask whether gauge R&R studies and calibration status are available. Measurement error can be as large as the variation you want to model.
Measure how often the root cause reads "operator error" or "retrain operator". A high share means the corrective-action labels say little.
Ask for the defect-code history and a mapping table if codes were renamed, merged or split.
Confirm each measurement is tied to the drawing revision in force when it was taken, since spec limits change between revisions.
Review sample images for faces, badges, screens and customer logos or markings on parts.
For medical-device or pharmaceutical plants, ask how complaint records containing patient information are kept out of scope.
Raise permitted-use limits early. A manufacturer may restrict uses that could benefit its direct competitors, which affects scope and price.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can I license real manufacturing quality and inspection data for AI training?

Yes. Quality records can be licensed when the manufacturer holds the rights, which is clearest for its own products and processes. SourceX looks for plants whose processes, materials and record types fit your spec, separates records each plant can license itself from those tied to customer designs, and fixes permitted use in a written license. Whether matching records turn up depends on which manufacturers take part, so supply is not guaranteed.

Can a contract manufacturer license records about parts it makes for customers?

Only within what its customer agreements allow. Drawings, specifications and often the inspection plan for a customer's part are confidential to that customer, so records tied to them need the customer's authorization or have drawing-derived content removed. Records about the plant's own processes, equipment and internal defects are usually easier to clear.

Do defect images come labeled?

They carry the labels the plant recorded: defect code, severity, disposition and sometimes a marked region. Those labels were applied to make production decisions, not to train models, so they vary between inspectors, shifts and years. If you need boxes or masks under your own taxonomy, say so in the request so relabeling can be scoped.

Is SPC and measurement data included, or only reports?

That depends on the plant's systems. Plants running an SPC or inspection database hold measured values per characteristic, with spec limits and gauge IDs, that export cleanly as tables. Others keep measurements only inside PDF or spreadsheet reports, which need extraction first. Each candidate dataset's manifest states which form they take.

Can inspection results be linked to machine and process data?

Sometimes. Where a plant's MES or process historian records settings and sensor readings by machine, lot and time, inspection results can be joined to them, pairing process conditions with the quality outcome for predictive-quality models. How fine the join can be depends on traceability: serial-level tracking supports part-by-part matching, while lot-level tracking supports only coarser joins.

Can quality records from regulated industries be licensed?

They can be, with extra review. Automotive, aerospace, medical-device and pharmaceutical plants keep detailed nonconformance and corrective-action records because their standards and regulators require them. Complaint records can contain patient information, and aerospace or defense work can involve export-controlled technical data, so those categories are reviewed separately and left out unless they can be cleared.

Field service and maintenance work orders
Work orders tracing symptom, diagnosis, parts and fix, with asset histories and photos
CAD and PCB engineering files with revision history
Native CAD and ECAD design files with revisions, change orders and BOMs
Supply chain and logistics operations records
Linked orders, shipments, tracking events, exceptions and freight documents from real operations
First-person video of skilled manual work
First-person video of skilled workers doing real tasks at partner businesses
Human feedback and QA-scored work
Work items with scores, verdicts and corrections from the people who reviewed them

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.