TellseedCreate a campaign
How it worksUse casesGuidesAgent setupPricingSign inCreate a campaign

AI training data

Human data collection for AI training and evaluation

Turn a data requirement into a campaign people can answer. Collect text, labels, corrections and ratings, then review the responses before they become training or evaluation data.

Last updated 3 October 2026

Have a task and an audience?

Create a free workspace and connect your agent, or follow the training-data guide.

For teams who know what data they need

Tellseed is a campaign tool for collecting human input. An AI agent can prepare the task through MCP, while contributors answer on a branded web page without a Tellseed account. You provide the audience and decide which responses are suitable for your model.

Start with a small text-based project: classify support requests, collect domain-specific answers, ask experts to correct a response, or evaluate an assistant against a rubric. The same product also handles ordinary surveys, customer feedback and research questionnaires.

Choose the human input your model needs

The task determines the answer format. Write the instructions and acceptance criteria before inviting contributors.

  • Text classification: show an example and collect one or several allowed labels using answer buttons.
  • Expert demonstrations: show a task and collect a written answer or correction for later supervised fine-tuning preparation.
  • Human model evaluation: collect ratings or structured judgements against a defined rubric.
  • Preference feedback: place two responses in the item context and ask for A, B, Tie or Neither. This uses ordinary answer controls; there is no dedicated pairwise comparison interface or automatic model-position randomisation.

From a brief to a reviewed export

Describe the task to your connected agent, supply the items or ask for an explicitly labelled synthetic pilot, and specify the response format. Ten items shown one at a time is the starting point. Larger pools, shorter rounds and collection ordering can be configured separately.

Preview the contributor page, approve the exact campaign version and publish. Invite your contributors, inspect coverage and review their responses. Keep the public collection link separate from private owner and reviewer links.

Exports include raw and reviewed JSONL, optional CSV, campaign version references and a manifest with counts, exclusions and file hashes. Scenario or contributor grouping can keep related records together when assigning train, development and test splits. Inspect the actual distribution, especially for small datasets.

Connect collection to your data pipeline

Use n8n to create a campaign from supplied items, publish after owner approval and export reviewed data. Signed webhooks announce new answers, completed questionnaire rounds, finalized reviews and ready exports. These deterministic workflows do not require another AI model call.

Keep original answers and corrected review targets separate in your destination. Deduplicate events by ID and retrieve current review status before using a target. For a training run, freeze a reviewed export, verify its file hashes and record the snapshot ID with your dataset version.

What you arrange separately

Tellseed does not recruit a participant panel, pay contributors, train models or guarantee that a dataset improves a model. You arrange recruitment, permissions for the source material and the intended use of contributions, your review process and your training pipeline.

Voice recording, audio transcription collection, video annotation, bounding boxes and token-span annotation are not current collection formats. For those tasks, use a specialist workflow. Do not treat written responses as a speech dataset.

A reviewed Tellseed export is not automatically a fine-tuning provider’s input format. Map the fields, inspect quality and document the dataset before training. Keep a separate evaluation set that has not influenced your training choices.

Start with one task and a small pilot

Try the website chat for a simple questionnaire, or create a hosted Free workspace and connect your agent for the full campaign and export workflow. A website trial lasts 24 hours and has a basic results preview. Check the current plan limits before collecting a larger dataset.

An answer is one submitted response to one item. For example, 100 contributors answering 10 items generate 1,000 answers. That describes collection volume, not 1,000 independently validated training examples.