All learning paths Learn See Try Make Build the examples it learns from What this lesson is about Two hundred careful examples beat twenty thousand careless ones. Write the careful ones yourself.
You have one week before a small model must answer shop questions in Hindi. How do you spend it?
A Download twenty thousand question-answer pairs from the web and hand the file over. B Write two hundred pairs quickly, so the week is spent on the count rather than on reading. C Two hundred careful examples beat twenty thousand careless ones.
Check my recall Why this matters A model copies what the examples show, including the tone and the mistakes. Unread examples teach whatever happens to be inside them. Writing two hundred yourself is slow, and it is the only part of this work that decides the result.
Build a small set of teaching examples you have read, with some kept aside for testing.
A model copies what the examples show, including the tone and the mistakes. Unread examples teach whatever happens to be inside them. Writing two hundred yourself is slow, and it is the only part of this work that decides the result.
Two hundred careful examples beat twenty thousand careless ones. Every example shows exactly the behaviour you want, including the refusals. Keep some examples aside from the start, so you have an honest test later. Task Build teaching examples for a small model that answers shop questions in Hindi.
Weak approach Downloads twenty thousand pairs scraped from the web and opens the first twenty. The whole file goes to the training script, and the rest is never read.
Stronger approach Writes two hundred pairs by hand, including ten where the right answer is a refusal. Keeps forty aside, unseen.
Why the stronger approach works The weak set teaches whatever was scraped. The strong one teaches the wanted behaviour and keeps an honest test.
Try a changed situation Your tuned model answers a question it should have refused. Your examples contained no refusals. What do you add?
Make something yourself Write fifty teaching examples for one job, plus ten you keep aside and never show. Put the rule you followed above them, in one or two lines.
Lesson 6 of 8 on Open models on a small budget next Tune a model without a big machine Part 2 ends in a work sample you can send
Read the primary or official source: Hugging Face TRL · dataset formats