Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one

Home›Services›Education & assessment›AI & data preparation

Research Data Preparation

AI and Data Preparation Services

Pubrica provides AI and data preparation services for researchers, institutions, publishers and AI development teams. Our data preparation specialists take raw, unstructured research data and deliver a cleaned, pre-processed, annotated and documented dataset that is ready for building, validating and deploying a model. Each delivery includes a decision log, a data dictionary and a bias and balance report, so the dataset can be reproduced by someone outside the project.

  • Clean, consistent and organised datasets: errors, duplicates, inconsistencies and missing values resolved, with every decision recorded
  • Accurate annotation and labelling for structured and unstructured data, to the standard required for supervised learning
  • Noise, bias and inconsistency reduced, with any remaining imbalance reported in full
  • Standardised metadata and documentation, so the dataset can be deposited, cited, reused and reproduced
  • Ethical data handling, with data governance, consent and provenance documented as part of the preparation
Scope of Data Preparation Included, excluded and client decisions

Cleaning, structuring and normalisationIncluded

Annotation and labellingIncluded

Bias and imbalance reportingIncluded

Metadata, documentation and provenanceIncluded

Building or training the modelNot included

Defining what counts as a labelClient decision

Raw → cleaned → annotated → validated → model-ready

Included in every data preparation order
  • Every cleaning decision logged
  • Inter-annotator agreement measured
  • Class imbalance reported
  • Provenance and consent recorded
  • A data dictionary for your records
3–10 business daysBy package; large datasets quoted separately
Measured annotation qualityInter-annotator agreement reported
Documented and reproducibleData dictionary and decision log delivered
ISO 9001:2015Quality management certified · cert. H3CR9TCU20251212INDQM10N

Service Scope

Six Components of Every AI Data Preparation Project

The scope is written on your order. Decisions on what counts as a valid label and what the model is for remain yours throughout.

Data Cleaning with Recorded Decisions We resolve errors, duplicates, inconsistencies and missing values and record every cleaning rule, so that each decision can be reconstructed and the dataset can be reproduced.

From the decision logDropped 1,204 rows with missing outcome · imputed age with the column meanRows with missing outcome retained and flagged; missingness is not random here and dropping them removes 31% of one site. Age imputed by site median, with an indicator column, and both choices logged.

Measured Annotation and Labelling We annotate structured and unstructured data to the standard required for supervised learning and measure agreement between annotators.

From the annotation report1,840 records, two annotators, blind. Cohen’s κ = 0.71 overall; 0.44 on the two ambiguous classes. Those 212 records were adjudicated by a third annotator, and the guideline was rewritten before the remaining 12,000 were done.

Structuring and normalisation We convert unstructured and raw data into consistent, documented formats that fit your pipeline, so that sources are combined only where they are compatible.

From the schema noteDates in 4 formats · sex coded M/F, 1/2, and free text · units mixedOne schema, one code list per field, units converted with the original retained, and a data dictionary that says what every column means and how it was derived.

Feature preparation and optimisation We transform, scale and select variables to improve training, fit transformations on the training split only and document them so the features can be reproduced on new data.

From the feature noteScaling fitted on the training split only and applied to the others — fitting on the whole dataset leaks information from the test set and flatters the model. 14 derived features documented with their formulas.

Bias, Imbalance and Noise Reporting We name and quantify the remaining problems in the data, such as class imbalance and skewed distributions, because a model inherits any imbalance in its training data.

From the bias report“Data balanced by oversampling the minority class.”Class balance 94:6 before, reported as such. Oversampling was applied inside cross-validation folds only; applied beforehand it inflates accuracy by roughly nine points and none of it is real. Site and age distributions reported alongside.

Documentation, governance and provenance We prepare standardised metadata, a data dictionary, a record of data sources and consent, and the governance documentation needed to deposit, share or publish the dataset.

From the dataset documentationData dictionary, 63 fields · decision log, 41 entries · provenance and consent per source · licence and reuse conditions · README written for users outside the project.

The samples shown are illustrative and anonymised. Hover over a sample to replay it.

Our Services

Six Types of AI and Data Preparation Work

These services are priced as packages because they depend on each other. For example, data must be harmonised before it is annotated.

Data cleaning and preprocessingWe resolve errors, duplicates, inconsistencies and missing values and record every cleaning rule.
Annotation and labellingAccurate annotation and labelling of structured and unstructured data for supervised machine learning, with a written guideline and measured agreement between annotators.
Structuring and normalisationWe convert raw and unstructured data into consistent, documented formats that combine safely in an AI pipeline or analytics workflow.
Feature preparation and optimisationWe transform, scale and select variables to improve training, fit them on the training split only and document them so they can be reproduced on new data.
Validation and quality assuranceWe check the prepared dataset for accuracy, consistency, leakage between splits and readiness, and report any remaining problems.
Compliance, ethics and data governanceWe document provenance, consent and licensing and work within the regulations that apply to your data, such as GDPR, HIPAA or your institution’s rules.

Data types we prepare: Text, tabular, image, audio and video data; research and experimental datasets; survey, clinical and observational data; and AI training and validation sets. Service exclusions: we do not build, train or tune your model, and we do not decide what a label means. Both decisions remain with you.

Who We Serve

Clients We Support with AI and Data Preparation

Requirements differ by client in the governance regime that applies and in what the dataset must satisfy, whether a journal, a repository, an ethics committee or a production system.

Academic researchers and data scientistsClean, annotated, AI-ready research datasets that support reproducible analysis, with the documentation reviewers and replication studies require.
Universities and research institutionsDatasets prepared and documented for AI-based research, archiving and cross-disciplinary reuse.
Publishers and research platformsDatasets and supplementary material prepared for publication and for deposit in open-access repositories, with consistent standards across submissions.
AI developers and industry teamsStructured, labelled and optimised data for training and validation, with leakage checks that reduce gaps between development and deployment.
Non-profits and government agenciesEthically prepared, compliant datasets supporting data-driven policy, research and programme evaluation.
Corporate and analytics teamsOperational and business data cleaned, harmonised and documented so analytics and automation work from a single agreed source.

Compliance and Standards

Seven Data Preparation Standards and Guidelines

Data preparation must meet research reporting guidelines, data-sharing practice and the data protection regime that applies to your data.

FAIRprinciples

FAIR data principles (Findable, Accessible, Interoperable, Reusable)

The principles against which deposited datasets are assessed, and the basis of our metadata and documentation work.

ICMJE Recommendations (January 2026)

Data availability statements and the declarations that go with a dataset published alongside a paper.

TRIPOD+AI, 2024

TRIPOD+AI (BMJ, 2024)

The reporting guideline for clinical prediction models, which includes specific questions on how the data were prepared.

C-AI2020

CONSORT-AI and SPIRIT-AI (Nature Medicine, 2020)

The reporting guideline for trials involving an AI intervention, which requires data handling to be reported.

COPE (Committee on Publication Ethics)

Publication ethics, including the position on data and on declaring how AI tools were used.

GDPRand equivalents

Data protection and governance (GDPR, HIPAA and your institution’s rules)

Where data concern people, the applicable regulations determine what preparation is permitted. We work within your regime and state clearly what we cannot do.

ISO9001:2015

Quality management (cert. H3CR9TCU20251212INDQM10N)

This certification covers how our work is run, checked and recorded. It is not a claim about the data itself.

Responsible Use of AI

Our Approach to AI Use and Disclosure

AI tools support specific, checkable tasks. Editorial judgement, final wording and sign-off remain with Pubrica subject specialists.

How We Use AI ToolsWe use AI tools to find and group issues at scale, such as inconsistent terminology, reference formatting and style deviations, which a specialist then reviews and resolves.
Where We Do Not Use AIWe do not use AI tools to write or rewrite scientific content, make editorial, ethical or integrity decisions, assess research quality or sign off any deliverable.
Your AI Use DeclarationYour declaration should name the tool, the stage at which it was used and the human oversight applied. Journals, funders, regulators and the ICMJE January 2026 update now expect this, and an undeclared tool found later is treated as an integrity matter.

Why We Do Not Publish AI Accuracy Figures

Accuracy figures for AI tools are usually reported by the companies that sell them.

Pubrica therefore does not quote its own accuracy figures, as they would carry the same conflict of interest. Instead, we tell you which tool was used on your project, at which stage, and what a specialist changed afterwards.

Methodology

How Our AI Data Preparation Process Works

We profile the dataset before any cleaning begins, so that every later step is based on a clear description of the data.

Step one

Initial assessment

We profile the dataset against your research objective, intended use and technical requirements, including missing rates, duplicates, coding inconsistencies and class balance.

Step two

Cleaning and structuring

The dataset cleaned, organised and standardised for accuracy, consistency and usability, with each rule recorded so the same result can be produced again.

Step three

Annotation and preparation

We carry out annotation, labelling and feature preparation aligned to your objective. The annotation guideline is written before labelling starts, and agreement is measured once a first batch is complete.

Step four

Optimisation

We improve clarity, balance and relevance while preserving the integrity of the original data. All transformations are documented and reversible.

Step five

Review and quality assurance

We check the dataset for accuracy, consistency and readiness, including whether any step could have leaked information between training and test splits.

Step six

Delivery and Documentation

You receive the prepared dataset, the data dictionary, the decision log, the bias and balance report and a README for users outside the project.

Request a Free Data Preparation Sample

We return the sample cleaned, structured and annotated, with a decision log, a data dictionary for those fields and a profile of issues in the rest of the dataset, such as missing rates, duplicates, class balance and inconsistent coding. The sample is delivered within two business days at no charge.

What We Need From You
The dataA representative sample of the dataset
The objectiveThe intended purpose of the model
The labelsYour label definitions, if available
The provenanceData sources and consent conditions
The targetPublication, deposit or production use
The deadlineYour timeline and dependencies

Please send a representative sample rather than your cleanest file. A quote based on the cleanest part of a dataset will not reflect the work needed on the rest.

Our Experts

Your AI Data Preparation Team

A named team of subject specialists stays with your project from scoping to delivery. Your project coordinator introduces each team member by name before work begins.

Subject matter expertClinical or scientific expert matched to your topic

Reviews the scientific content for accuracy and confirms that each conclusion is supported by the evidence. Each expert is matched to your field.

OwnsScientific accuracy and interpretation
Reviews againstThe applicable curriculum or subject standard
FieldMatched to your topic
Specialist editorReviews educational content against the applicable standard

Reviews content for accuracy, pedagogy and conformance with the applicable curriculum, assessment or accessibility standard.

OwnsContent accuracy and conformance
Reviews againstThe applicable curriculum, assessment or accessibility standard
OutputReviewed, annotated files
Second reviewerIndependent quality check before delivery

Checks the finished work independently against the agreed scope and the applicable checklist. Findings are logged and resolved before delivery.

OwnsIndependent quality control
Reviews againstThe agreed scope and checklist
Signed off byProject lead
Project coordinatorYour single point of contact

Manages the timeline, keeps you informed at each stage and coordinates the team so that questions are answered and deadlines are met.

OwnsTimeline and communication
Reviews againstYour brief and deadline
OutputDelivery and follow-up

Sample Work

Sample Clinical Dataset: As Received and As Delivered

Step through the six findings one at a time. Three of them are issues that make a model appear more accurate than it is, which are often not detected until deployment.

6 changes, each explained

As received. 18,400 records, 63 fields, four sites, three export formats.

  • 1,204 records missing the outcome — concentrated in one site
  • Sex coded as M/F, 1/2, and free text
  • Dates in four formats, two ambiguous
  • Class balance 94:6, unreported
  • 412 near-duplicate records
  • No data dictionary, no provenance record

Delivered. Cleaned, annotated and documented, with remaining issues reported.

  • Retained and flagged — missingness is not random; dropping them removes 31% of site C
  • One code list, original values retained in a shadow column
  • ISO 8601 throughout; 14 ambiguous dates raised as queries, not guessed
  • 94:6 reported — resampling inside CV folds only
  • 388 merged with a documented rule; 24 raised as queries
  • Data dictionary, decision log, provenance and consent per source

Record counts and field names have been changed. Agreement statistics and imbalance figures are taken from the delivered report on the original project.

Redacted extract · counts and fields changedAdvanced package · clinical tabular data

Benefits

Benefits of AI Data Preparation

Each benefit can be checked in your delivered dataset. None is a claim about model performance, which depends on more than the data.

01 Reliable Performance MetricsLeakage and resampling in the wrong order are common causes of models that perform well in development and fail in deployment. Both are preparation decisions, and neither is visible in the metrics.
02 Bias Reported TransparentlyA model inherits any imbalance in its training data. Reporting the imbalance allows you to decide how to address it.
03 Reproducible DatasetsThe decision log and data dictionary allow the same dataset to be rebuilt from the raw source and defended in review.
04 Measured Annotation QualityInter-annotator agreement gives annotation quality a measurable figure. Where agreement is low, the guideline is revised before the remaining records are labelled.
05 Ready to Deposit, Cite and ReuseFAIR metadata, provenance and licensing are prepared alongside the data, so publication or deposit does not require a separate project.
06 Ambiguities Raised as QueriesWhere the data cannot resolve a question, we ask you. We do not insert estimated values that cannot be traced.

Packages

AI Data Preparation Packages

Each package is priced on data volume, the extent of annotation required and the governance work the data needs. We prepare a free data profile first and quote on that basis.

Basic

Basic data preparation

Researchers and analysts who need a clean, structured dataset

Quoted per projectPriced on volume

Delivery 3–5 business days

  • ✓Cleaning and error correction
  • ✓Structuring and formatting
  • ✓Duplicates and inconsistencies resolved
  • ✓Standard metadata prepared
  • ✓Initial quality checks and a decision log

Best for research datasets, pilot projects and academic studies.

View a sample Request a quote
Advanced

Advanced AI data optimisation

Projects needing structured, annotated and optimised data

Quoted per projectPriced on annotation volume

Delivery 5–7 business days

  • ✓Everything in Basic
  • ✓Advanced preprocessing
  • ✓Annotation and labelling with a written guideline
  • ✓Inter-annotator agreement measured and reported
  • ✓Feature preparation and optimisation
  • ✓Bias, imbalance and leakage checks

Best for machine learning models, analytics projects and journal-linked datasets.

View a sample Request a quote
Premium

Data engineering and compliance

Universities, enterprises and large-scale AI initiatives

Quoted per projectPriced on scale and governance

Delivery 7–10 business days

  • ✓Everything in Advanced
  • ✓Large-scale dataset preparation
  • ✓Ethical AI and data governance alignment
  • ✓Provenance, consent and licensing documented per source
  • ✓Multi-format delivery
  • ✓Expert QA review and a validation report

Best for enterprise systems, funded research and high-stakes deployments.

View a sample Request a quote

Send 500 rows or 50 records and we will prepare them free of charge, with a profile of issues in the rest of the dataset. Large-scale and ongoing work is quoted against a schedule.

Quality Assurance

Our Service Commitment

Our service commitment, its limits and the remedy we offer are set out below.

Our quality commitment

Every Dataset Reproducible from the Raw Source

The deliverable includes the decision log, the data dictionary and the bias report, which allow others to rebuild, check and defend the dataset.

01

Every Decision Documented

We record what was dropped, imputed, merged or recoded, and why, so each cleaning rule can be reconstructed.

02

Annotation Agreement Measured

Annotation quality is reported as an agreement statistic, with the adjudication record behind it.

03

Remaining Issues Reported

Class imbalance, missingness patterns and distribution skew are stated in the report.

04

Free Rework or a Refund

If the work does not meet the level on your order, we revise it at no charge within the agreed scope, or refund it if you prefer.

Client Stories

“They turned raw datasets into clean, model-ready resources. The change in model performance was immediate — and lower than before, which turned out to be the honest number.”
JohnsonData scientist

USA

FAQs

AI and Data Preparation FAQs

If your question is not answered here, ask an expert — no quote is required.

Poor or unstructured data limits the performance of any model. Data preparation determines accuracy, reliability and reproducibility.

Several common preparation errors also make a model appear more accurate than it is, including leakage between splits, resampling before cross-validation and imputation fitted on the whole dataset. None of these appear in the metrics.

Text, tabular, image, audio and video data; research and experimental datasets; survey, clinical and observational data; and AI training and validation sets.

Clinical and survey data usually require the most governance work, and we include this in the scope from the start.

We follow the FAIR principles for documentation and metadata, the reporting guideline that applies to your work, such as TRIPOD+AI for clinical prediction or CONSORT-AI for trials with an AI intervention, and the data protection regime that applies to your data.

Your institution’s data standards and your funder’s requirements also apply, and the order states which ones are in scope.

Yes. We agree a written guideline before labelling starts and measure agreement between annotators once a first batch is complete.

Where agreement is low, the guideline is usually revised, and it is more efficient to do this on the first batch than at the end of the project.

Yes. Preprocessing, annotation and structuring depend on the model type, the domain and the intended use, and a dataset prepared for one model may not suit another.

Please tell us the objective when you send the sample, so that the preparation is designed for that use.

No. This service ends with a dataset that is ready for model training. We keep data preparation separate from model building so that the team preparing the data does not also report model performance.

Biostatistics and predictive analytics are separate services for analysis and modelling, and are quoted separately.

Related Services

Related Data Collection and Analysis Services

This service ends with a model-ready dataset. The services below cover study design and data collection before preparation, and analysis afterwards.

Request a Free Data Preparation Sample

We return the sample cleaned, structured and annotated, with a decision log, a data dictionary and a profile of issues in the rest of the dataset, such as missing rates, duplicates and class balance. Delivered within two business days at no charge.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.