进行模型精调前，需要准备符合模型类别和精调方法要求的训练集与验证集。数据集采用 JSONL 格式，每行包含一个完整的 JSON 对象。

支持模型

图片生成模型支持 SFT，具体模型 ID 和版本以创建精调任务时控制台可选项为准。

SFT 精调

图片生成模型总体上分为两类数据，一类是文生图 (t2i) 数据，特点为输入只有 prompt。另一类是编辑（edit）数据，即图生图，特点是输入是文本 + 单图/多图。两类数据均采用相同的格式，以下是完整的格式示例、说明及字段定义。

每条样本必须满足以下要求：

messages 中恰好包含 2 条消息，第一条为 user，第二条为 assistant，不支持 system 消息或多轮对话。

user.content 必须且只能包含 1 个 text 类型内容作为生成指令，可以包含 0 个或多个 image_url。

assistant.content 必须且只能包含 1 个 image_url，不能包含 text 类型内容。

格式示例

样本格式：为JSONL文件（ JSON Lines，是一种轻量级的文本文件格式，核心规则每一行对应一个独立的、合法的 JSON 对象），需确保单个对话样本独占一行，示例如下。

{"messages":[{"role":"user","content":[{"type":"text","text":"生成一只可爱的小狗照片"}]},{"role":"assistant","content":[{"type":"image_url","image_url":{"url":"tos://bucketname/image/dog.jpeg"}}], "loss_weight": 1.0}],"extra":{"sample_id":"sample-001","metadata":{"source":"user"}}}​
{"messages":[{"role":"user","content":[{"type":"text","text":"生成一只可爱的小狗照片"}]},{"role":"assistant","content":[{"type":"image_url","image_url":{"url":"data:image/jpeg;base64,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"}}], "loss_weight": 1.0}]}​
{"messages":[{"role":"user","content":[{"type":"text","text":"生成一只可爱的小狗照片"}]},{"role":"assistant","content":[{"type":"image_url","image_url":{"url":"file:./image/dog.jpeg"}}], "loss_weight": 1.0}]}​
{"messages":[{"role":"user","content":[{"type":"text","text":"把图2中的猫换成图1中的狗"},{"type":"image_url","image_url":{"url":"tos://bucketname/image/dog.jpeg"}},{"type":"image_url","image_url":{"url":"tos://bucketname/image/cat.jpeg"}}]},{"role":"assistant","content":[{"type":"image_url","image_url":{"url":"tos://bucketname/image/dog_in_forest.jpeg"}}], "loss_weight": 1.0}]}​

格式说明

为便于展示各个字段关系，分别从JSONL 格式文件中取出一条t2i类别的数据和一条edit类别的数据，具体如下：

字段说明

每行一条 JSON 格式数据，核心字段如下：

字段名					类型	是否必填	说明
messages			list	是	对话列表，必须恰好包含 2 条消息，不支持多轮对话或 system 消息。
	role		str	是	第一条消息必须为 user，第二条消息必须为 assistant。 user：支持图文混排 assistant：仅支持图片内容
	content		list	是	对话内容：当 role 为 user：必须且只能包含 1 个文本指令，可以包含 0 个或多个参考图片当 role 为 assistant：必须且只能包含 1 张图片，不能包含文本内容
		type	str	是	子内容类型，可选text（文本）或image_url（图片）
		text	str	条件必选 type为text时	文本内容，不能为空字符串
		image_url	dict	条件必选 text为image_url时	图片信息，包含url字段（图片地址）。参考：使用限制
	loss_weight		float	否	仅 assistant 角色支持配置该字段，默认 1.0，取值范围 [0.0,1.0]。user 角色不支持该字段。
extra			dict	否	样本顶层的用户自定义扩展字段。extra中的 key 必须为 string 类型，value 类型不限制。

图片生成模型的图片传入限制，请参见使用限制。

火山方舟