Concept Contrast Fable

Fine-tuning and instruction fine-tuning: one is major surgery, the other is changing the way you converse

Conclusion first:When teams discuss model optimization, fine-tuning and instruction tuning are often treated as the same operation. Fine-tuning uses domain data to continue training all or some parameters of a model, changing what knowledge the model has learned; instruction tuning teaches the model to follow the instruction formats given by humans, changing how the model understands and responds to tasks. Use the wrong type, and costs cannot be calculated accurately, nor can results be measured correctly.

When teams discuss model optimization, fine-tuning and instruction tuning are often treated as the same operation. Fine-tuning uses domain data to continue training all or some parameters of a model, changing what knowledge the model has learned; instruction tuning teaches the model to follow the instruction formats given by humans, changing how the model understands and responds to tasks. Use the wrong type, and costs cannot be calculated accurately, nor can results be measured correctly.

Fable

Two types of apprentice transformation in the carpentry workshop

The old carpentry workshop had two kinds of apprentices. The first kind already knew how to make furniture, but every time they received a custom order—say, a Ming-style curio cabinet—they had to go to the client's home to measure dimensions, understand the style, and try to reproduce it. The master found this too inefficient, so he picked a few apprentices with solid fundamentals and had them practice making large quantities of Ming-style furniture in the workshop. After intensive practice, these apprentices could produce drawings directly in the workshop for new orders without going on site.

The second type of apprentice had a different problem. Their craftsmanship was not bad, but they could not understand customers' verbal descriptions. When a customer said, "The cabinet should feel lighter," the apprentice did not know whether to make the legs thinner or the panels thinner. The master gave this group of apprentices a translation training instead of continuing to teach specific furniture-making methods: break down the customer's colloquial language into action instructions—select material, set proportions, adjust decoration. What the apprentices learned was the conversion from "lighter" to "reduce the leg diameter by two fen."

Once an old regular ordered a set of study furniture. The first type of apprentice produced a curio cabinet with beautiful proportions and precise lines—even the old masters nodded. The second type of apprentice produced work that was technically competent but wrong in character—the "understated elegance" the client asked for became "don't add anything" on the drawings, and the finished piece looked plain. The two types of rework addressed shortcomings at completely different levels.

The workshop owner later summed up a rule: if apprentices simply cannot make a certain type of furniture, have them focus on practicing how that type of furniture is made; if apprentices can understand drawings but not spoken instructions, specifically train their ability to convert speech into drawings. The two kinds of training cannot replace each other.

Once an urgent order tripped up quite a few people. The client said, "Make a cabinet that can store tea cakes," and a new apprentice who had previously received specialized training in tea cabinets built a fully functional tea cake cabinet. But he had not received instruction-understanding training, and when several other clients in the same batch asked him to "make something in a style young people like," he had no idea where to start changing.

Later, the workshop added a round of basic training for all apprentices: no longer techniques for specific furniture, but learning to break any customer description into executable steps. When these apprentices received new tasks, at least they knew where to start—even if they didn't do it well enough, they wouldn't be completely stuck.

The workshop's teaching manual later added a line: specialized techniques make you more precise on a certain type of task, while instruction understanding keeps you from freezing when you encounter an unfamiliar new task. The two paths differ completely in training data, evaluation methods and investment costs, yet they had been written on the same line of the project management sheet for far too long.

Fine-tuning

On top of a pretrained model, continue training the model parameters with data from a specific domain or task so the model acquires new knowledge or capabilities.

Instruction fine-tuning

Train a model with instruction-answer pairs so it learns to understand task descriptions and output in the required format; this is a specific subclass of fine-tuning.

Story Mapping

A large amount of Ming-style furniture practice: fine-tuning, using domain-specific data to make the model master a certain type of task
Learn to translate 'lightness' into shrinking the leg diameter by two parts: instruction fine-tuning, letting the model understand colloquial task descriptions and act on them
The apprentice makes curio cabinets excellently but cannot understand style requirements: only domain fine-tuning, lacking instruction alignment
Basic translation training added for all apprentices: instruction fine-tuning as a general capability patch

Applying it to an Agent project

When evaluating a model, first distinguish whether what you need to solve is insufficient domain knowledge or failure to understand the task. If domain knowledge is insufficient, prepare industry corpora for fine-tuning; if the task is not understood, collect instruction-answer pairs for instruction fine-tuning. The data volume, GPU cost, and evaluation metrics of the two types of training are completely different. Running them mixed together both costs more money and leaves you unable to see where the bottleneck is.

Fine-tuning Instruction fine-tuning