Prepare a reviewed training-data export

Raw and reviewed files answer different questions

The raw export preserves collected responses with their review state. It is useful for auditing collection and understanding what was excluded. The reviewed export contains finalized, accepted examples, including reviewer corrections where those settled the answer.

A reviewed file can be empty even when people have answered. Check the manifest exclusions: examples may be unreviewed, unresolved, not accepted, awaiting a second review or from a superseded version. Do not call raw data reviewed simply to make an export non-empty.

Tellseed supports JSONL and optional CSV exports. Full exports require access to full results; an unpaid website trial only exposes a basic preview. Use the current pricing page to check the workspace route and limits.

• Current plans and access

Understand the record fields you will map

In reviewed JSONL, recordId identifies the record, scenarioExternalKey identifies the source item and lineageGroupId connects related versions of an item. campaignVersion and campaignVersionDigest identify the campaign wording and configuration used for the answer.

input.context holds the item context; input.brief holds the shared brief when the campaign includes it in exports. target contains the typed response. targetSource tells you whether it came from the submission or a reviewer correction. provenance contains submission and review references, and split identifies the assigned dataset split.

For a text-classification task, a single-choice target contains a selected value. For a written answer, a free-text target contains text. Keep response kinds explicit in your conversion code; do not concatenate every field into model input.

Contributor email is excluded from training exports. That does not make every exported record anonymous: identifiers and free-text content can still be identifying. Review and minimise the fields you actually distribute or train on.

Choose a split strategy that matches the leakage risk

Tellseed defaults to grouping by scenario lineage when assigning train, development and test splits. Related versions of a scenario stay together. Contributor grouping instead keeps one contributor’s responses together. The none strategy assigns records individually and does not provide either grouping protection.

The default ratios are 80/10/10, with a reproducible seed. Assignment is by hashed groups, so small datasets may not match those proportions and can have an empty development or test split. Inspect the manifest counts and empty-split list before evaluation.

The manifest records leakage checks for the selected grouping. This does not detect every semantic duplicate or protect both scenario and contributor identity at once. Near-duplicates, shared source documents and subject overlap need additional checks in your dataset preparation.

Save the manifest with the actual files

The manifest records the snapshot, campaign versions, policies, counts, exclusions, split assignment details and SHA-256 hashes for exported files. Verify the files against those hashes before using them in a pipeline and retain the preparation script version alongside them.

A snapshot records a particular export selection. New submissions do not silently appear in the old snapshot. Contributor erasure can withdraw an affected export, so do not assume old downloaded copies will disappear automatically; maintain your own deletion and redistribution process.

Convert deliberately for your training workflow

For supervised fine-tuning, map the approved input and accepted answer into the exact schema your training framework expects. For preference tuning, join the preference to both original candidate responses and apply your policy for ties, missing answers and rejected responses.

Tellseed does not train a model, upload your dataset to Hugging Face or automatically emit every provider’s fine-tuning format. Test the conversion with a small validated fixture and keep the original export so errors can be traced back.

Before release, document purpose, source permissions, contributor instructions, review rules, transformations, languages, known coverage gaps and permitted uses. Use a dataset card to make that context available to the next person who touches the files.

• Dataset card template
• Hugging Face dataset card documentation