Managed SFT data

Text conversations

Each records item contains messages ending in nonempty assistant text. Use the quickstart request. The platform freezes the original submitted content. Tokenization, shifted targets and optimization are managed by the recipe. The loss is computed on the final assistant message only (loss_on: last_assistant_turn on Qwen/Qwen3.8-27B; see options); system, user and earlier assistant messages are context, so a multi-turn conversation trains its last answer.

Images

A user message may contain {type: "text", text: "..."} and {type: "image", image: "data:image/png;base64,..."} parts. Only bounded, decodable, single-frame PNG/JPEG/WebP inline images are accepted. The assistant target remains text. Remote URLs and caller-local image paths are not accepted. Limits include 1 MiB/image, four images/record, 4 MiB total image bytes, and bounded pixel counts; the public request limit also applies. Do not log private samples or inline bytes.

Other recipes

Other configured models may use their own frozen bounds. The pev27 recipe uses imported data snapshots and named presets, not arbitrary inline records. Read discovery/schema before selecting it.

Mixed tasks

Balance examples and evaluate per task. Managed SFT does not imply RL, custom loss or arbitrary user-defined training code support.