Examples

Sanitized examples showing how teams structure datasets for fine-tuning runs.

These examples are sanitized. Validation runs before payment; if validation fails, you won't be charged.

Customer Support Tone Adapter

Use case:: Consistent, empathetic replies across support channels with clear escalation boundaries.
Dataset format:: JSONL with chat-format records ({"messages": [...]}), each containing a customer message and approved response, tagged by scenario.
Validation checks:: JSONL validity, required fields and roles, dataset size, token cap, record count, and max line length.
Expected outcome:: More consistent tone and fewer unnecessary escalations.

Use case:: Accurate answers to repetitive product questions with technical precision.
Dataset format:: JSONL prompt/answer records, each grounded to canonical product documentation sources.
Validation checks:: JSONL validity, required fields, dataset size and token caps per tier, record count, and max line length.
Expected outcome:: Higher first-response accuracy in self-serve channels.

Use case:: Reliable extraction from semi-structured text into a stable schema for automation.
Dataset format:: JSONL with instruction-format records ({"instruction": ..., "output": ...}) aligned to a single extraction schema.
Validation checks:: JSONL validity, supported instruction/output structure, dataset size and token caps per tier, record count, and per-record line-length limits.
Expected outcome:: Cleaner downstream automation with less manual cleanup.