Training file format
Valcairn trains on JSON lines: a .jsonl file with one JSON object per line. Each line takes one of three shapes. Lines in any other shape are skipped, and the import check names each one.
Chat messages
A messages list with system, user and assistant turns. The last turn must be the assistant answer that is trained, after a non-empty user turn. OpenAI chat fine-tuning files use this shape.
{"messages":[{"role":"system","content":"You answer questions about Acme orders in one or two sentences."},{"role":"user","content":"Where is my order?"},{"role":"assistant","content":"You can follow it from the link in your confirmation email."}]}Download chat-example.jsonl (6 rows)
Prompt and completion
A non-empty prompt and the completion to learn. Each row becomes one user question and one assistant answer.
{"prompt":"Classify the ticket: \"My parcel arrived broken.\"","completion":"damaged_item"}Download completion-example.jsonl (6 rows)
Preference pair
A prompt with a chosen and a rejected answer, for preference tuning. The two answers must differ; identical pairs are skipped.
{"prompt":"Reply to: \"My parcel is late.\"","chosen":"Sorry about the wait. It is on its way and should arrive tomorrow.","rejected":"Parcels are sometimes late."}Download preference-example.jsonl (6 rows)
Before a job can start
A file needs at least 5 distinct prompts: about one in ten is held out to measure the training, and the rest must still be enough to train on. Alternate answers to the same prompt count once and always stay on the same side. The training runner reads only files ending in .jsonl, at most 2 GB per file and 20 GB per training job.
Not read today: tool calls and tool results, image or audio parts, and per-message weights (every assistant turn in a kept chat is trained).