Where fine-tuning fits

What you'll learn

• Explain why pre-training and fine-tuning, which share one training objective, teach such different things.
• Compare a web-scale pre-training corpus with a curated fine-tuning set by size, labels and who runs each stage.
• Choose between the text-only and the question-and-answer storage formats for a training example.

Lesson 1 of T14.1 showed a base model continuing Tell me how to train my dog to sit with more questions, and a chat model with the same architecture answering it. This page places those two models on one timeline. A Larkspur engineer who has read that comparison sometimes concludes the house style needs a model trained from scratch on the handbook. The comparison below shows why that is the wrong stage.

Both stages train the same way: show the model text and adjust the weights so it predicts each next token better. What differs is the data each stage feeds in:

Figure 1: One objective, two datasets. Pre-training produces the base model; fine-tuning reshapes it.

The notebook prints both kinds side by side. The first rows it streams from C4, a web-crawl corpus, are scraped pages: whatever text was on the page, in whatever order. The fine-tuning file, lamini_docs.jsonl, has 1,400 rows, each a question and an answer about one company's product. Pre-training gives the model language and general knowledge from a huge, unlabelled corpus. Fine-tuning shows it, in a small labelled one, what a good response looks like:

Table 1: The two stages compared. Only the data and the operator differ; the objective is the same.

The Larkspur house style is a format and a tone, which puts it in the right-hand column. A few hundred curated answers in that style, trained on top of an existing base model, do the job that a from-scratch run on the handbook could not: the handbook alone is far too small to teach language, and it contains policy text, not examples of good support answers.

Key distinction: pre-training decides what the model knows how to say; fine-tuning decides which of those things it says when a request arrives.

Once the rows exist, the next decision is how to store them.

The notebook formats the same lamini_docs rows two ways, and both are common:

• Text only: one text field holding the whole example, ### Question: and ### Answer: included. Trainers that take a single text column, such as AutoTrain in lesson 7 of T14.1, expect this.
• Question and answer: the prompt in one field (### Question: up to ### Answer:) and the answer in another. Keeping them apart lets you score only the answer at evaluation time, and mask the prompt out of the loss if you choose.

Both are written to JSON Lines, one example per line, which the dataset loader in lesson 4 of T14.1 reads. Here is one Larkspur example in the second form, field by field:

Table 2: A Larkspur example in the question-and-answer form (the 30-day window comes from the Larkspur handbook's data export page).

In the text-only form the same example is a single text value: the question field and the answer field joined. That join is where the next lesson starts, because the exact characters between them, and after the answer, are what the model learns to reproduce.

Picking a form is a choice about the later steps:

Table 3: Which storage form each later step needs.

The third row is the common case. Keeping the fields apart costs nothing, and a join is one line in the data script.

The code for this lesson is in these notebooks. Colab runs one in your browser: install the packages its first cell imports and set the API keys it reads. To run every notebook as written, clone the repository and use its uv environment.

• Pre-training and fine-tuning use the same next-token objective. Only the data differs: web-scale raw text for language and knowledge, a few hundred to a few thousand labelled pairs for a format, task or tone.
• Model labs pre-train once per model. You fine-tune whenever the behaviour needs to change.
• Store examples as JSON Lines, either as one text field for trainers that read a single text column or as separate prompt and answer fields so you can score only the answer.