BergLabs.

Multimodal data annotation & evaluation

Multimodal data, annotated and evaluated with review on every batch.

BergFlow controls intake, specialist review, multi-pass QA, reviewer accountability and lineage from receipt to delivery.

Send one batch See BergFlow
Data pipelineReceipt · Review · QA · Delivery
  1. VideoRaw videoCapture sessions
  2. Text + audioTranscriptsPrompts, calls
  3. SensorSensor framesLidar, IMU
Active sampleActive sample frameA video frame of a street scene with one car outlined by a bounding box labelled car, confidence 0.98.car 0.98

Specialist review · lineage kept

  1. QATwo gatesGround truth, sign-off
  2. Model teamDeliveredWith lineage
Illustrative sample. Label, confidence and gates vary by programme and are agreed in the SLA.
  • Who it is for

    AI and ML teams training or evaluating models.

  • What you get

    Reviewed, versioned data with lineage and QA evidence.

  • How it is priced

    By scope, modality, volume and quality bar.

  • 96 to 98%Review accuracy, typical
  • 2+QA gates per batch
  • 1,400+Trained specialists

Review accuracy is above 90% as a rule, typically 96 to 98%, agreed per project in the SLA.

How it runs

From receipt to delivery, on one governed record.

Three controls that keep a dataset usable after it leaves us.

  1. 01Structured intake

    Every batch arrives with its schema, guidelines, ground truth and acceptance criteria attached, and every item gets a governed record on receipt.

  2. 02Multi-pass QA

    Specialists label, reviewers check, and sampled QC runs against ground truth. Each pass is recorded with the person who made it.

  3. 03Feedback-loop ready

    Model errors and disagreements come back as new tasks, so the next batch fixes what the last evaluation found.

Modalities

Eight modalities, one quality standard.

The same intake, review and QA gates apply to every modality.

  • Text curationFiltering, rewriting and ranking text to a rubric.
  • Image reviewBoxes, polygons, keypoints and attributes, reviewed.
  • Video segmentationFrame-accurate segments with continuity checks.
  • Audio transcriptionTranscripts, speakers and intents, multilingual.
  • 3D point-cloudLidar and depth labelled in 3D, synced to camera.
  • Preference dataPairwise and ranked judgments for model tuning.
  • Golden setsGround-truth sets that QA is measured against.
  • Reviewer QASecond-pass review with reviewer accountability.

What we deliver

Four ways teams use the service.

  1. 01

    Sourcing to spec

    Data collected to an agreed specification, with provenance recorded for every item. Licensing and consent terms are agreed per programme.

    Discuss this service
    1. 01SpecifySpec agreed
    2. 02CheckSample review
    3. 03DeliverProvenance kept
    1. 01Coverage, balance and exclusion rules are agreed before collection.
    2. 02Samples are reviewed against the spec before volume is collected.
    3. 03Items arrive with provenance on the record.
  2. 02

    Curation

    Deduplication, filtering and balancing, so a dataset represents the cases your model will meet.

    Discuss this service
    1. 01FilterFiltered
    2. 02ReviewSpecialist call
    3. 03BalanceBalanced
    1. 01Duplicates, near-duplicates and out-of-scope items are removed.
    2. 02Edge cases are judged by a specialist, not a script.
    3. 03The final set is balanced to the agreed distribution.
  3. 03

    Annotation and review

    Labels across eight modalities, reviewed by specialists, with QA evidence on every batch.

    Discuss this service
    1. 01LabelGuideline
    2. 02ReviewTwo gates
    3. 03DeliverWith lineage
    1. 01Specialists label against the guideline and golden set.
    2. 02Two QA gates per batch: sampled QC, then sign-off.
    3. 03Batches ship with lineage and QA reports.
  4. 04

    Evaluation sets

    Golden sets, preference data and held-out evaluation sets that your team can trust as a yardstick.

    Discuss this service
    1. 01DesignCoverage agreed
    2. 02AdjudicateAdjudicated
    3. 03LockVersioned
    1. 01Coverage and difficulty are agreed with your model team.
    2. 02Disagreements are resolved by senior reviewers.
    3. 03The set is versioned and locked before use.

Have a batch that needs a second look?

Send a sample. We grade it against your guidelines and come back with scope, SLA and a price.

Send a sample batch

Questions

Questions about multimodal data annotation and evaluation

Scope, quality and who does the review.

Data programmes

What does data sourcing and curating include?

Collecting data to an agreed specification, filtering and deduplicating it, annotating and reviewing it, and delivering it with lineage and QA evidence.

How is quality measured?

Every batch passes at least two QA gates and reports review accuracy against ground truth. Accuracy is above 90% as a rule, typically 96 to 98%, agreed per project in the SLA.

Which modalities do you cover?

Text curation, image review, video segmentation, audio transcription, 3D point clouds, preference data, golden sets and reviewer QA.

Who does the review?

Trained specialists vetted through Quinn, with reviewer accountability recorded on every item in BergFlow.

Can you work from our guidelines?

Yes. Task templates, guidelines, ground truth and SOPs are attached at intake, and disputed items carry their history to the next reviewer.

Send one batch through the queue.

We grade it, run it through BergFlow and return the QA evidence with the data.

Send one batch Start a pilot