Data Services

Usable multimodal dataand judgement, for AI

Art imagery, audio, engineering drawings and sensor signals — non-standard data plus expert judgement, turned into data you can train and evaluate on.

Scope

Art imagery, audio, industrial drawings and sensor time series — datasets, non-standard annotation and evaluation benchmarks.

Send a data briefSee the MUB benchmark
arXiv:2610.22989MUB · Music Understanding Benchmark

Partners

  • ByteDance

Data services

From raw material to dataset, every step leaves a record.

Dataset construction

Sample boundaries, class taxonomy and splits, delivered ready to train.

  • Filtering and deduplication
  • Train / validation / test splits
  • Data cards and versions

Sourcing and rights

Public collections, partnerships and own capture, with source and rights logged per item.

  • Source-chain logging
  • Rights scope
  • Sensitive-content screening

Cleaning and structuring

One format and resolution, with work, artist, period and medium completed.

  • Format and resolution standards
  • Metadata completion
  • Duplicate and corrupt samples

Evaluation benchmarks

“Is it right?” broken into reviewable questions and scoring rules.

  • Task and scoring rules
  • Reference answers
  • Versioned evaluation sets

From drawings to waveforms

CAD files, PCB layouts, simulation contours, sensor signals — hand engineering data to frontier models as-is.

CAD drawings

Read views, dimensions and tolerances; extract key sizes, generate BOMs, and diff revisions.

PCB designs

Parse netlists and footprints, cross-check part numbers, and surface rule-based review findings.

CAE results

Interpret stress and flow contours, map parameter effects, and draft structured simulation reports.

Industrial time series

Analyze sensor waveforms and machine logs, locate anomalies, and outline likely root causes.

Annotation solutions

The hard part is not tagging. It is exclusivity, edge cases and keeping batches on one definition.

Fine-grained visual annotation

Subject, composition, colour, light, brushwork and material, item by item.

Art-historical semantics

Period, school, medium, subject and attribution, by annotators with an art-history background.

Multimodal alignment

Image regions, captions and structured fields anchored to each other.

Custom taxonomies

Hierarchies, value sets and exclusivity rules, with a handbook.

Example label schema

Work attributes

ArtistPeriodSchoolMediumSizeProvenance

Visual elements

CompositionColourLightBrushworkMaterialSpace

Semantics & affect

SubjectNarrativeSymbolismMoodStyle intensity

Rights & compliance

Rights statusUsage scopeAttributionSensitivity

Schemas are trimmed to the task; each group ships with definitions, edge cases and exclusivity rules.

Quality

Quality is not sampled in at the end.It grows out of the process.

Handbook first

Definitions, edge cases and scoring rules written before annotation starts.

Two-pass review

A second annotator reviews key batches; disagreements go to arbitration.

Traceable versions

Every delivery carries a handbook version, manifest and change log.

Delivery workflow

Five stages, each with a confirmation point.

  1. 01

    Scope

    Task, target model, sample range, acceptance criteria

  2. 02

    Sampling & rights

    Sample by distribution; log source and rights

  3. 03

    Annotation

    Annotate by the handbook; flag disputes

  4. 04

    Quality review

    Independent second pass; disputes arbitrated

  5. 05

    Structured delivery

    JSONL / COCO / custom schema

ResearcharXiv:2610.22989·cs.SD·2026-10-16

MUBMusic Understanding Benchmark

500 identity-masked recordings, four tasks: tempo, key, time signature, structure. Models see audio and the instruction only; the data composition, masking protocol and scoring scripts are public.

Authors · Kun Liu · Lyra Tao

Read the paperGitHub
MUB composition and masking flow · 1080pH.264 · 15s · 500 KB
500
Recordings

MUB-00001 – MUB-00500

4,676
Structural intervals

5–22 per track

24
Key combinations

268 major / 232 minor

47–180
BPM range

Median 101

The paper describes this as an evaluation resource still being validated: recording-level clearance is incomplete and no model comparison has run. Code is MIT; audio is not distributed.

Read the full text

Where it fits

Multimodal training

Image–text pairs, region captions, VQA

Generation quality

One prompt set and rule set to compare models

Cultural heritage

Structured cataloguing and metadata

Platform governance

Style, subject and sensitive-content labels

Start with the data problem you actually have

Tell us the scale, target model and format; we will come back with a plan.

Send a data briefView model services

JSONL / COCO / custom schema