Skip to main content
A file is source material that Oumi does not train on directly, such as a product manual, a policy document, or a knowledge base article. Synthesis jobs use your files to generate conversations about your own content, so a model learns your products and policies instead of generic answers.

​
WHAT FILES ARE FOR

A dataset holds examples that Oumi trains on or evaluates. A file holds the facts those examples should be about. In Generate Synthetic Data from Scratch, add a file under Input Sources > Documents. Oumi splits the document into chunks that the job can write conversations about, such as questions on each section of a manual. The Files page accepts PDF, TXT, Markdown, and HTML documents. It also keeps tables, such as CSV or Parquet files, that don’t match a dataset shape. A file you upload that does match a dataset shape becomes a dataset instead.