Dataset Formats

Data shapes expected by each training method

MethodRequired shape
SFTprompt/completion, text, or chat messages
RAFTprompt rows for generation and verification
DPO/ORPO/RMprompt, chosen, rejected
GRPOprompt rows plus a verifier
VLMimage path or image payload, prompt, answer
Audioaudio path or decoded audio field, transcript or label
Reasoningquestion and answer, optionally reasoning trace
Agenticmessages, tool schema, tool call, final answer

Local JSONL paths can be selected in Train with Custom local file. Built-in dataset short names use the same registry as the CLI.