AI Data
How to scope a first dataset
The most common mistake is committing to volume before the specification has been tested. A small end-to-end pass costs less and tells you more.
3 min read
Services
The data your AI needs — specified, collected, labelled and validated.
From first-person human demonstrations to annotated audio, video and text, every service runs on the same operating standard: a written specification, trained people, human review and structured delivery.
One operating standard
The same six gates, whichever data type you buy.
One quality process for every service, so mixed corpora stay consistent.
- 01
Collection
Capture or sourcing against a defined specification, so the input is already in scope.
- 02
Guidelines
A written definition of correct, with worked examples and explicit edge-case handling.
- 03
Annotation
Trained annotators working to that guideline, escalating ambiguity rather than guessing.
- 04
QA
Structural and consistency checks across the batch, not just within single items.
- 05
Human Validation
Review by a second person against the acceptance definition.
- 06
Final Dataset
Structured delivery with the guideline and review notes attached.
01
Egocentric Data Collection
First-person human demonstrations for Physical AI — specified, captured on calibrated rigs, labelled to the sub-action and delivered with provenance.
What you receive
- Calibrated RGB or stereo video, with IMU clock-aligned to frames
- Depth, gaze and 3D hand pose where the rig tier provides them
- Temporal segmentation to sub-action, with outcome markers
- Provenance and consent manifest per episode
- LeRobot, RLDS, MCAP or raw + manifest
Useful to share with us
- The task families or skills your policy needs to learn
- Target environments and any object or grip constraints
- Preferred modalities and delivery format
- Your acceptance criteria, if already defined
Why it matters and how we work
The problem
VLA and world-model teams are limited less by architecture than by diverse, well-specified human demonstration. Commodity footage arrives unsynchronised, repetitive and without consent records, and much of it never reaches a training run.
What Mvaak does
We translate your skill catalogue into capture briefs, plan environments, objects, grips and failure modes for spread, and collect on calibrated rigs with worker-level consent — then run multi-stage QA and annotation before delivery.
02
Data Collection
Custom collection programmes run by trained, supervised field teams — planned against a written specification rather than assembled from whatever exists.
What you receive
- Collection brief and coverage matrix agreed before capture
- Raw data with task, environment and operator metadata
- Consent and provenance records
- Coverage report against the agreed specification
Useful to share with us
- What the model needs to learn and where it currently fails
- Required data types, environments and conditions
- Approximate volume and timeline, if known
Why it matters and how we work
The problem
Where coverage is thin, waiting for the right examples to appear in production data is slow and unreliable. Unplanned collection produces volume without the spread a model needs.
What Mvaak does
We plan collection against the cases your model needs — environments, participants, conditions and edge cases — and run it with trained collectors, field supervision and consent captured at source.
03
Audio Data
Transcription, segmentation, speaker labelling and audio event classification — to conventions that are decided once and applied consistently.
What you receive
- Transcripts to a written convention
- Timestamped segmentation and speaker diarisation
- Audio event labels
- Guideline and review notes attached
Useful to share with us
- Sample recordings and their source
- Languages and dialects required
- Existing transcription conventions, if any
Why it matters and how we work
The problem
A recording is not yet a dataset. Inconsistent transcription conventions and loose timestamps are among the most common reasons accurate transcripts still train poorly.
What Mvaak does
We transcribe, segment and label audio you already hold, and collect prompted audio where accents, phrasings or acoustic conditions are under-represented.
04
Video Annotation
Bounding boxes, tracking, classification and temporal segmentation — with occlusion and identity decisions written down before work starts.
What you receive
- Boxes, tracks, frame and clip labels
- Activity, event and temporal segmentation labels
- Consistency review across sequences
- Schema and guideline delivered with the data
Useful to share with us
- Representative footage samples
- Your label set or the behaviours you need detected
- Target output schema
Why it matters and how we work
The problem
A box drawn correctly on one frame and inconsistently on the next teaches a tracker the wrong thing. Most disagreement happens at occlusion and when objects leave and re-enter frame.
What Mvaak does
We annotate footage you hold against a specification written with your team, define the label schema where one does not yet exist, and check consistency across sequences rather than frame by frame.
05
Text Annotation
Classification, entities, intent, sentiment and span relations — with disagreement treated as a signal that the schema needs tightening.
What you receive
- Labelled corpus in your schema
- Agreement analysis on the pilot sample
- Human validation of edge cases
- Guideline with worked examples
Useful to share with us
- Sample texts
- Your label set or task definition
- Volume and language requirements
Why it matters and how we work
The problem
Text annotation looks simple until two annotators disagree about the same sentence. Unresolved definitions reach the model as noise.
What Mvaak does
We label documents, transcripts, tickets and conversations against a schema defined with your team, using a small first pass to expose disagreement before volume begins.
06
Data Quality & Validation
Field checks, automated integrity scoring, human review against a rubric, diversity scoring and blind acceptance sampling.
What you receive
- Written acceptance rubric
- Per-batch QA results and rejection reasons
- Diversity report against your catalogue or class list
- Recommendations for re-collection or re-annotation
Useful to share with us
- The dataset or a representative sample
- How the data will be used in training
- Any existing guidelines or acceptance criteria
Why it matters and how we work
The problem
Quality problems found after delivery cost the most to fix. Without a written acceptance rubric, nobody can say whether a batch passed.
What Mvaak does
We run data through defined quality gates — integrity checks, rubric-based human review, diversity scoring and blind re-grading — and report what passed, what failed and why.
07
Annotation & Dataset Rescue
Already have footage? We run unqualified archives through the same QA gates, salvage what is trainable and label it.
What you receive
- QA triage report on the archive
- Labelled, salvaged subset
- Format migration to LeRobot, RLDS or MCAP
- Independent quality audit, where requested
Useful to share with us
- Description and size of the existing data
- Current format and label schema
- What the data is intended to train
Why it matters and how we work
The problem
Many teams hold archives they cannot use — unqualified footage, inconsistent labels, or data delivered in the wrong format by another vendor.
What Mvaak does
We triage what you hold, salvage the trainable portion, add label depth, migrate formats and can run an independent quality audit — so you know what you actually own before you buy more.
Work with Mvaak
Four ways to start.
Start small with a pilot, then scale on the same quality standard.
30 days · Fixed scope
Start herePilot Sprint
A small, fast first engagement designed to show whether the approach works on your task.
- One task family, one domain
- Skill specification written with your team
- 100–250 accepted hours delivered
Quarterly · Committed volume
Managed Corpus Programme
Rolling capture against your skill catalogue, with a predictable delivery rhythm.
- Rolling capture against your skill catalogue
- Monthly delivery cadence and gap reporting
- Volume-committed pricing with quality SLAs
Ongoing · Exclusive
Dedicated Capture Cell
A team trained only on your specification, using your rigs and your capture protocol.
- Collectors trained only on your specification
- Your rigs and your capture protocol
- Embedded QA lead reporting to your team
Project-based
Annotation & Rescue
Label depth, QA triage and format migration for data you already hold.
- Label depth on data you already hold
- QA triage of unqualified archives
- Format migration to LeRobot / RLDS / MCAP
Commercials. Priced per accepted hour by rig tier and annotation depth. Rejected hours are not invoiced. A detailed rate card is issued on request after scoping.
Related reading
All articlesDataset Quality
What actually makes training data usable
Volume is the easiest property to measure and the least predictive of whether a dataset improves a model. Here is what tends to matter instead.
3 min read
Data Annotation
Annotation guidelines are the real deliverable
Two annotators disagreeing about the same example is not a training problem. It is a definitions problem, and it is far cheaper to solve before collection than after.
3 min read
Start a project
Need a custom dataset?
If your requirement does not map cleanly onto one service, it is usually a combination. Describe the model and what it gets wrong — we’ll tell you what the data would have to look like.