Data Hub & Datasets

Infere provides two dataset surfaces: the Data Hub for curating collections of production requests (sampling, branching, deduplication, export) and Evaluation Datasets for the datasets that evaluation runs execute against.

• Data Hub - create datasets from random samples of production traffic (by token, model, status), from templates (training, evaluation, few-shot, error analysis), by merging, or as editable branches. Export as JSONL (OpenAI Batch), JSON (Anthropic), or CSV.
• Evaluation Datasets - examples with train/dev/test partitions, human annotation workspaces (Pass/Fail/Skip verdicts), and frozen versions so runs stay reproducible. Import production traces as examples with one click.
• Dataset analytics & quality - per-dataset cost, latency, success rate, completeness, and data-health checks (duplicates, missing cost data, schema validation).