All learning paths Learn See Try Make Decide when tuning is worth it What this lesson is about Most jobs need better instructions, not a tuned model. Learn to spot the few that do need tuning.
A small model puts only half your Marathi complaints in the right field. Your teammate wants to start a week of training. What now?
A Start the week of training now, since half right means the model has not learnt the job. B Rewrite the instruction once, and if half is still wrong, start the week of training. C Most jobs that look like tuning need better instructions and better examples.
Check my recall Why this matters The Hugging Face course says to consider tuning only when you need more than prompting can give. Most gaps close with clearer instructions and a few examples inside the request, which cost an afternoon. Tuning costs a week and a machine, so it goes last.
Decide with two measured numbers whether a job needs tuning or only a better request.
The Hugging Face course says to consider tuning only when you need more than prompting can give. Most gaps close with clearer instructions and a few examples inside the request, which cost an afternoon. Tuning costs a week and a machine, so it goes last.
Most jobs that look like tuning need better instructions and better examples. Try in order: clearer instructions, examples in the request, then tuning. Tuning pays when the job has a fixed shape and a few hundred careful examples. Task Decide whether to tune a small model for sorting Marathi complaints into three fields.
Weak approach Starts a week of tuning because plain instructions put only about half the complaints in the right field. The instruction itself is never rewritten and no examples are added.
Stronger approach Rewrites the instructions, then puts six examples in the request. Scores the same fifty each time and writes both numbers down.
Why the stronger approach works The weak choice spends a week to learn what an afternoon answers. The strong one has two measured numbers before spending anything.
Try a changed situation A colleague wants to tune a model because its answers are too long. Nothing else is wrong with them. What do you tell him?
Make something yourself Pick one job. Write what you tried first, what it changed on your own fifty examples, and whether you would tune. Give the two numbers, not an impression.
Lesson 5 of 8 on Open models on a small budget next Build the examples it learns from Part 2 ends in a work sample you can send
Read the primary or official source: Hugging Face LLM course · supervised fine-tuning