S5-07 · 基础 · 概念实验 S5-07 · Foundation · Concept Lab

微调 · 换上下文和改模型,不是一回事 Fine-tuning · Changing Context Is Not Changing the Model

same need → prompt / retrieval / fine-tuning → compare parameter state

微调(Fine-tuning)是在已有模型参数上继续训练,使其适应特定任务、领域或行为。它真的改变参数,因此和提示、RAG 的推理时控制必须分开。 Fine-tuning continues training from existing model parameters to adapt a task, domain, or behavior. It truly changes parameters and must be distinguished from inference-time prompting or RAG.

01 · 先从日常问题开始 01 · Start with an everyday question

员工临时看一张操作卡,和参加两周培训后形成习惯——这两种改变会留在同一个地方吗? An employee reads a one-off job aid versus trains for two weeks. Do those changes live in the same place?

02 同一需求走三条轨,只看哪里动了参数 Route One Need Three Ways and Watch Which Changes Parameters

选择路线并调样本数 choose route + examples
参数变化 Parameter change
toy Δθ
上线成本 Upfront cost
relative units
回滚难度 Rollback effort
relative level

03 三条路线改变的是不同层 The Three Routes Change Different Layers

把刚才的变化拆开 Map the change to parts
01 · PROMPT

提示改输入 Prompt changes input

每次请求带上任务、约束或示例,不改基础参数。 Each request carries tasks, constraints, or examples without changing base parameters.

02 · RAG

检索改上下文 Retrieval changes context

在推理时加入外部证据,知识可更新,参数保持不变。 External evidence enters at inference time; knowledge can update while parameters stay fixed.

03 · FT

微调改参数 Fine-tuning changes parameters

用适配数据继续训练,行为变化进入新版本权重。 Continued training on adaptation data moves behavior into a new parameter version.

04 · OPS

版本与回滚 Version and rollback

改参数后必须保存基线、检查点和可切换版本。 Parameter changes require a baseline, checkpoints, and switchable versions.

04 最小判别:有没有发生参数更新 Minimal Test: Did Parameters Update?

fine-tune: θ′ = θ − ηg
prompt/RAG: θ′ = θ
fine-tune: θ′ = θ − ηg
三条轨使用同一需求。提示和检索改变当前输入;只有微调根据适配样本计算更新并产生新参数版本。 All routes receive the same need. Prompting and retrieval change current input; only fine-tuning computes updates from adaptation examples and creates a new parameter version.
同一业务需求 Same business need 固定比较起点 fix the comparison
选择控制层 Choose control layer 输入、上下文或参数 input, context, or parameters
比较成本回滚 Compare cost and rollback 按变化位置治理 govern where change lives

05 边界实验:零样本也叫微调 Boundary Test: Call Zero Examples Fine-tuning

主动制造失败 Create a failure

选择微调,却不给任何适配样本 Choose fine-tuning with no adaptation examples

没有样本就没有更新方向;即使按钮写着“微调”,参数变化仍为 0。 Without examples there is no update direction. A button labeled fine-tune still produces zero parameter change.

尚未执行。先在上方改变主变量。 Not run yet. Change the main variable above first.
把文档塞进提示就是微调。 Putting documents in a prompt is fine-tuning. → ✓ 那是改变上下文,基础参数不变。 That changes context; base parameters stay fixed.
微调后无需保留旧版本。 The old version is unnecessary after fine-tuning. → ✓ 参数变化必须支持检查点与回滚。 Parameter changes require checkpoints and rollback.
任何需求都应优先微调。 Fine-tune every new need first. → ✓ 先比较提示、检索、成本与更新频率。 Compare prompting, retrieval, cost, and update frequency first.
06 · 一句话带走 06 · One line to keep

提示改输入,检索改上下文,微调才继续训练并改变参数;选路线时必须同时比较成本、更新频率与回滚责任。 Prompting changes input, retrieval changes context, and fine-tuning continues training to change parameters. Route selection must include cost, update cadence, and rollback responsibility.