Creative Data Lab Start a project
Datasets for AI  ·  San Francisco, California

Data, made
with intent.

We design, build and curate the training and evaluation datasets that models actually learn from.

Start a project → How we work
raw curated 37.77° N · 122.42° W
01 — What we do

Most AI problems are data problems in disguise.

Create

New data, from scratch

When the data you need doesn't exist, we build it: expert-written examples, carefully captured images and video, and synthetic data checked by people.

Curate

Better data from what you have

Deduplication, filtering, rebalancing, relabeling and decontamination: we turn a noisy corpus into a set you can train on with confidence.

Evaluate

Tests your model hasn't seen

Held-out benchmarks and rubric-scored eval sets that measure the capability you care about, not just leaderboard numbers.

02 — Domains

Where we go deep.

Language

Post-training

SFT, preference and reasoning data that shapes how a model behaves, not just what it knows.

  • Expert-written instructions and responses
  • Pairwise preferences with rationales
  • Verified multi-step reasoning traces
Measurement

Evals & benchmarks

Private, contamination-checked test sets built around your product's real failure modes.

  • Rubrics and grader calibration
  • Adversarial and long-tail cases
  • Versioned splits that never leak
Perception

Vision & multimodal

Image, video and depth annotation for models that need to see the physical world.

  • Detection, segmentation, density maps
  • Video temporal grounding
  • Image–text and video–text pairs
03 — Process

From a fuzzy goal to a versioned dataset.

  1. Week 0Scope

    We start from the model behavior you want, then work backward to a data spec, schema and acceptance criteria.

  2. PilotPrototype

    A small pilot batch to check the spec against your training loop before we scale anything.

  3. BuildProduce

    Sourcing, creation and annotation by vetted contributors, with automated checks running throughout.

  4. QAVerify

    Layered review, agreement scoring, and model-assisted audits to find errors before your model does.

  5. ShipDeliver

    Versioned splits, a full dataset card, and provenance for every row, in the format your pipeline expects.

04 — Standards

Every dataset ships with its receipts.

  • Provenance. Where every example came from, and who touched it.
  • Rights. Licensed or consented sources only, documented per source.
  • Privacy. PII detection and scrubbing before anything leaves our pipeline.
  • Quality. Measured agreement and error rates, reported rather than assumed.
  • Hygiene. Deduplication and decontamination against public benchmarks.
dataset_card.yamlexample
name:        reasoning-prefs-v1
domain:      post-training / preference
format:      jsonl
splits:
  train:     —
  eval:      —   # held out, never published
fields:      [prompt, chosen, rejected, rationale]
labels:      [preference, error_type]
provenance:
  source:    expert-written, in-house
  consent:   documented
quality:
  review:    2-pass human + automated
  agreement: reported per label
pii_scrub:   true
decontam:    checked
license:     per client agreement
05 — Who we are

Creative Data Lab is founded by Srijan Bhushan, a data engineer who has built production data pipelines and hands-on AI projects, from local LLMs to video and vision models.

We stay small on purpose, so the people who scope your data are the people who build it.
06 — Contact

Tell us what your model needs to learn.