Data Intelligence × Human Experts

Representative build volumes by domain

Domain	Representative volume
Translation & language corpora	Three datasets of 1.5M Korean–English sentence pairs each (2020–2021), 2.54M sentences across five financial language pairs, 2.5M Korean–Chinese / Korean–Japanese patent and engineering parallel items
Speech & interpretation	2,017 hours of specialist Korean–English conference interpretation speech with 900K transcribed and translated sentences, plus 118 hours of defence-domain speech
Video & image multimodal	41,000 images with 410K descriptions for Korean-context video understanding, 52,463 K-Stock images with 1.15M captions, 3,900 educational videos with 12,078 question–answer pairs
Specialist domains	200,830 LLM training items for civil law and IP law, 300,240 labeled patents mapped to the national S&T classification, 400,452 multilingual tourism menu translations

The full list, with the year, role, and build volume of each dataset, is on the R&D achievements page. Volumes follow the detail published on AI Hub; because source, labeled, and evaluation data are counted separately for some datasets, summing every figure would double-count.

Dataset design and processing for your training goal

Multimodal dataset construction

We design and process advanced multimodal training data that combines video, audio, images, and text. The critical part is securing cross-modal alignment quality that single-modality datasets cannot provide.

Media-specific datasets

Broadcast subtitles, webtoon dialogue, image descriptions, multilingual translation, and custom metadata — datasets built on media intelligence that reflect the data shapes actually used in media operations.

Enterprise-grade refinement

Source data from specialized domains such as patents, law, environment (ESG), and engineering is refined to agreed training and delivery requirements, including noise identification and removal, schema normalization, and label taxonomy design.

Three-stage quality management

Training data requires more than collecting source material. We manage source quality, label definitions, consistency between workers, and delivery specifications. AI supports repetitive processing, with experts reviewing the results.

Stage	What is checked	Basis for the next stage
Design review	Training objective, data types, labels, and sample composition	Agreed instructions and sample results
Processing review	Errors and omissions in transcripts, translations, descriptions, and labels; source alignment	Corrected, reviewed data in consistent formats
Final dataset review	Agreed volume, format, and quality criteria	Final dataset incorporating review results

Dataset quality is checked against agreed criteria. The performance of a model trained on the data is evaluated separately. Error tolerances, review methods, and sampling scope are defined at the start.

Experience operating large data projects

Six consecutive years of NIA AI training-data work, from 2020 to 2025, covered media, language, education, law, finance, defence, and mobility. Consortium collaboration, workforce management, large-volume data handling, and security and infrastructure operations inform delivery.

Clients can follow design, processing, and review as one connected process. Specialist work includes agreeing on the expertise needed to review terminology and context, while keeping source materials, outputs, and corrections distinct.

Delivery criteria agreed before work starts

• Goal and scope: intended model and task, data types, languages, domains, and required volume.
• Samples and instructions: label definitions, exception handling, and formats checked on a small batch.
• Quality and review: inspection items, sample scope, error decisions, and rework criteria.
• Schedule and data handling: milestones, interim reviews, security requirements, and transfer methods.

See the complete R&D project register for dataset names, years, roles, and public build volumes.