词元之火

About us

Sparking creativity and value in everyone

01 — Model Services

Outside mainland China

Model servicesModel Services

One API for frontier models.

Available and billed outside mainland China only.

View model services
01

Unified access

OpenAI-compatible — swap the base_url and go.

02

Smart routing

Chosen in real time by latency, uptime and price.

03

Transparent billing

Tokens and cost logged for every call.

04

Service assurance

Uptime standards and support are ours to keep.

02 — Data Services

Datasets · Annotation · QA & acceptance

Data servicesData Services

We turn non-standard data, together with expert judgement, into structured datasets that can be trained on, evaluated and accepted.

View data services
01

Dataset building

From sample evaluation to volume delivery; reproducible samples, classes and splits.

02

Sourcing & rights

Source, licence scope and sensitive content logged per item; samples are separate from training rights.

03

Annotation schema

Multimodal label schemas and an 840-entry dictionary, extended per task with mutual-exclusion rules.

04

Quality review

Key fields independently reviewed twice; every delivery includes a quality report and error analysis.

Data assets

Samples to evaluate. Datasets to deliver.

A sample of delivered data; other modalities and industries can be scoped per project.

mechanical CAD files
≈28,000
annotated audio
202.5 h
structure segments
75,371
label dictionary entries
840
Mechanical 3D CAD

CAD

Mechanical 3D CAD

Real parts, moulds and tooling: about 11,000 native NX files (roughly half with CAM programs), about 4,200 native SolidWorks files, about 12,300 STEP files and about 4,300 2D drawings.

  • NX 2 – NX 2412
  • Full modelling history
  • CAM toolpaths
  • Drawing–model pairs
  • Product–mould pairs

≈28,000

files · 4 formats

Typical tasks

CAD generation and completion, drawing understanding, process planning, model retrieval

Non-exclusive, perpetual, worldwide licence for training, fine-tuning, testing and evaluation. Evaluation samples carry no training rights.

Music understanding

MUSIC

Music understanding

Segment-level annotation of real songs: structure, BPM, time signature, key, rhythm and melody notes, instrumentation, vocals, genre, mood, scene and language. Built from 3,327 annotation tasks, 202.5 hours of audio and 93 annotators.

  • Structure segments
  • BPM / meter / key
  • Instrumentation / vocals
  • Genre / mood / scene
  • Lyric alignment
  • 840-entry dictionary

1,430

reviewed final songs

Typical tasks

Structure analysis, lyric alignment, multi-label classification, audio model evaluation

Final songs passed QA and were de-duplicated per track; annotator identity and submission timestamps are removed on export.

Business partners

ByteDance

Quality & acceptance

Quality is not a promise.It is a report shipped with the data.

Every batch includes a quality report, error analysis and acceptance record. Metrics are calculated per batch; a batch that misses the bar is reworked in full.

Shipped with each batch

  • Structured dataset
  • Quality report
  • Error analysis
  • Data card & file manifest

Field completeness

Non-empty fields that pass the enum check / required fields

Per-field statistics

Dual-annotation agreement

Matching independent reviews / reviewed samples

Disagreement list

Interval validity

Intervals without overlap or gaps, covering start to end / all intervals

Automated check log

Dictionary hit rate

Labels resolved by the enum dictionary / all labels

Unresolved label list

First-pass acceptance

Batches accepted on first review / all batches

Batch trend

Error closure

Errors fixed and regression-checked / errors found

Error analysis report

Acceptance flow

  1. 01

    Sample pre-check

  2. 02

    Metric review

  3. 03

    Dispute & arbitration

  4. 04

    Revised delivery

Public benchmark

MUB · Music Understanding Benchmark

500 identity-masked recordings evaluating tempo, key, time signature and structure.

Sample evaluation and training licences are signed separately. Every delivery includes a source manifest, annotation manual and version log.

Business inquiry

Tell us what you need. We will reply by email.

Submissions are stored in our console and followed up by a person. For anything urgent, email works too.

0/500

Contact

Let's talk models and data.

outreach@tokenspark.uno
Headquarters
Shanghai, China