AI training data
Training data for AI models, labeled to your rubric
Need labeled data to train or evaluate a model? On trainee.digital you define the label set and the acceptance bar, set the price per unit, and a vetted annotation workforce produces the dataset. Escrow-backed: funds release only when you approve the batch.
Dataset work we produce
| Model need | Data we produce | Unit |
|---|---|---|
| Computer vision — segmentation | Per-image masks/regions with class labels from DICOM or standard image exports | per image |
| Computer vision — detection/tagging | Bounding boxes, tags and attributes on images, video frames or satellite tiles | per image / per tile |
| NLP — classification | Document- or clause-level labels with span offsets | per page |
| NLP — extraction | Key-value field extraction from scanned documents | per page |
| LLM alignment (RLHF) | Ranked preference pairs over prompt/response sets with rationale codes | per pair |
| Model evaluation | Human scoring of model outputs against your rubric; disagreement flagged for review | per item |
| Speech | Transcript correction and error tagging on audio clips | per clip |
How a dataset order runs
- Step 1
Define the rubric
You provide the label set, edge-case rules and the minimum accepted quality score. Ambiguous cases must be flagged with a note, not guessed — that rule is enforced in vetting.
- Step 2
Price and fund it
Volume × price per unit + 10% platform fee, shown before publishing. The amount is funded into escrow.
- Step 3
Annotation by vetted workers
Only accounts that passed the scored 50-unit assessment (70% bar, human review) can take the order.
- Step 4
Review, approve, release
You review the delivered batch. Funds release only on your approval; rejected batches are reworked or cancelled.
Delivery, formats, and iteration
- In: you keep hosting the source data and share access; typical inputs are JSONL prompt sets, image or DICOM exports, PDFs, and audio clips.
- Out: JSON or CSV label files per batch, plus notes on flagged units so you can fix the rubric before scaling.
- Iteration: start with a few hundred units, read the flagged notes, then scale. Rubric ambiguity is the usual reason a dataset disappoints, and it shows up in flags first.
- Evaluation sets: keep a held-out slice labeled by a second performer if you need inter-annotator agreement — request it in the order guidelines.
Escrow-backed milestone payments: how money moves
- You fund the order before work starts. Funds sit in escrow on the platform and are not paid to performers at that moment.
- Work is delivered per milestone. You review the delivered batch against the order guidelines.
- Funds release only when you approve the milestone. If you reject it with a reason, the batch is reworked or the milestone is cancelled and the escrowed amount stays with you.
- Performer balances become withdrawable after approval. Minimum withdrawal is $100; requests are processed manually with the payout details the performer submits.
- Disputes follow our Dispute Resolution process: email first, then a review of the submission, order requirements and platform records.
FAQ on training data
Can you guarantee accuracy?
We enforce a process, not a magic number: a 70% scored assessment for access, your minimum accepted quality score written into the order, and your approval gate before money moves. Accuracy on your data is measurable with a paid pilot batch.
How fast can a dataset be produced?
Set the deadline yourself in the order (current open orders run 4–10 days at 800–3,000 units). Throughput follows the per-unit rate you offer.
Do you train models for us?
No. We produce and QA labeled data; training and deployment stay on your side.
Is the data used for anything else?
No. Labels and order metadata are stored to run the marketplace — see the Privacy Policy.
Contact and who runs this
trainee.digital is operated by NGO "ElectronNick" (EDRPOU 4480186765007), registered in Odesa, Ukraine. Written questions about orders, assessments, payouts or invoices: support@trainee.digital. We reply within 1–2 business days. See also our Terms of Service, Privacy Policy and Dispute Resolution process.