Add examples
Examples are inputs, and the right answers when you have them. They show Facet what the task looks like in your app. You don’t need many: Facet writes more on its own.
Where to add them
- When you make a Model: your coding agent sends the inputs it found in your repo with the description, or a file you point it at. Facet uses them to draft the Model and its evals. Inputs without answers are fine: Facet writes the answers.
- From your agent, later: the
add_examplestool adds examples to one eval, andcreate_uploadsends a larger file as training data or hidden eval data. - On an eval in the web app: open it and choose Upload JSONL or CSV to add examples to that one statement. Turn on Hide these to keep them for checks only.
Formats
CSV with a header row. Facet reads the input from a column named input, prompt, text or question, and the answer from output, completion, expected, answer or label. Without an input column, the whole row is the input.
input,output
"I was charged twice this month",billing
"The export button does nothing",bug
JSON lines, one object per line, with the same field names. Any other object is taken whole as the input.
{"input": "I was charged twice this month", "output": "billing"}
{"input": "The export button does nothing", "output": "bug"}
A JSON array of the same objects works too, as does plain text with one input per line. See Limits for how many rows each way takes.
Answers with several fields. When a Model returns fields, the answer is the whole object: an object in JSON lines, or the same JSON as text in the CSV cell.
{"input": "I was charged twice this month", "output": {"label": "billing", "urgent": true}}
input,output
"I was charged twice this month","{""label"": ""billing"", ""urgent"": true}"
How many
- To start: 5 to 20 real inputs. More helps only if they show something new.
- Per eval: a few that show what it means, including the hard cases.
- Answers: optional. Without them Facet writes its own and checks them; with them, Facet matches yours.
Facet writes examples until each eval has about 200 to build from, and keeps some hidden to check with. You see every example it writes, and you can edit or remove any of them.
What to leave out
- Personal data you don’t need. Facet treats everything you upload as private, but a Model rarely needs real names or account numbers to learn the task.
- Examples that disagree with each other. If two tickets that read the same have different labels, Facet can’t know which one you mean.
Next steps
- Write good evals: what to say about your examples
- Privacy and data: who sees what you upload