Docs
Datasets
What dataset formats does Onrup accept and how are they validated?
Upload JSONL, CSV or Parquet in a conversation, instruction or preference shape. A format adapter converts it to the trainer-ready form. Validation runs before any compute is leased and reports counts, token-length percentiles, duplicates and truncation risk.
Shapes, not just formats
The file format — JSONL, CSV, Parquet — matters less than the shape of each record. Three shapes cover almost everything: a conversation with role-tagged turns, an instruction with an optional input and an output, and a preference record with a prompt plus a chosen and rejected response.
A format adapter converts whichever you have into the trainer-ready form for your chosen objective. That separation means changing the target model does not require rebuilding your data.
| Source shape | Suits | Converts to |
|---|---|---|
| Conversation, role-tagged turns | Multi-turn assistants | Chat format for supervised training |
| Instruction, input, output | Single-turn tasks | Chat format for supervised training |
| Prompt, chosen, rejected | Preference tuning | Preference pairs |
| Prompt only | Reinforcement learning | Prompt set for reward-scored generation |
| Tabular rows | Classification and extraction | Chat format via a column mapping |
Uploads are resumable
The API returns presigned URLs and the bytes go straight to storage. Nothing large passes through the API, which is why a multi-gigabyte upload does not time out and an interrupted one resumes from the last completed part rather than starting again.
Read the report properly
The validation report exists to stop you paying for a run that was doomed at upload. Four numbers deserve attention every time.
- Token length percentiles — the 95th percentile against your sequence length tells you the truncation risk before it happens.
- Truncation count — how many targets would be cut off. Anything above a per cent or so is worth fixing before training.
- Duplicate rate — exact and near duplicates inflate the weight of repeated content and contaminate held-out splits.
- Class balance — for classification, a category with thirty examples will be handled badly however good the aggregate looks.
Seeds and generation
A seed asset is a small, high-quality set intended to be expanded. Generation runs under a cost ceiling declared up front and stops at that ceiling rather than exceeding it, and every generated record carries provenance so it can be told apart from collected data later.
Generation is a tool for filling identified gaps, not for manufacturing a dataset from nothing. A model trained wholesale on another model’s output inherits its blind spots without the capacity to recover from them.
Frequently asked questions
How much data do I need?
For format and tone, a few hundred consistent examples make a visible difference. For task competence, low thousands. Consistency matters more than volume — contradictory targets teach the model that the task is ambiguous.
Is my data used to train anything else?
No. Datasets stay in your tenant’s storage and are used only for the runs you start. Publishing pushes weights and a model card, never data.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.