What is Distillation?
Knowledge distillation, told as a story about collecting a big parrot's answers and teaching a small parrot to copy them.
The big parrot is great. But it is slow, and every answer costs a heap of beans.For a job that asks thousands of times a day, that adds up. A small parrot would be nice — but a small one is not as good.
Raising a small parrot on the whole mountain of books again costs far too much.The reading school (pretraining) costs a million sacks of beans. You cannot reopen it for one job.
Distillation collects heaps of the big parrot's answers and teaches the small parrot to copy them.The big parrot is the teacher, the small one the student. It learns from the teacher's answer notes instead of books — far cheaper and faster than school.
Thousands of the big parrot's answer notes drill the small parrot.Step 3 is fine-tuning. Because the teacher's notes were not written by people, they are called synthetic data.
The student rarely passes the teacher, and copies the teacher's mistakes too.So always confirm with a test. Teaching from someone else's parrot raises terms-of-service issues. Sending only the easy jobs to the small parrot is routing; shrinking it another way is quantization.
Distillation = collect heaps of a big parrot's answers and teach a small parrot to copy them. Far cheaper than school, and good enough for our job. The student rarely passes the teacher and copies its mistakes.
Training a small student model on the outputs of a large teacher model, keeping task performance while cutting cost and latency. The teacher outputs are synthetic data, and the student training is fine-tuning (SFT). Training on another vendor's model outputs may violate its terms of service.
When grown-ups say it
- Knowledge distillation
- Teaching the small parrot to copy. The big parrot's answers train the small one.
- Teacher · student model
- The teacher parrot and the student parrot. The teacher is big and good; the student is small and fast.
- Synthetic data
- Answer notes a parrot wrote. A workbook made by a parrot, not people. → teaching a parrot with a parrot's book
- Small language model (SLM)
- The small parrot. Fits on a phone or laptop. Good at one job.
- Fine-tuning
- The drill on answer notes. How the student learns. → the short drill
- Versus quantization
- Teach a new bird vs lighten the same bird. Distillation raises another parrot; quantization rewrites the same parrot's numbers coarsely. → counting beans coarsely
- Routing
- Which parrot gets the job. Easy jobs to the student, hard ones to the teacher. → which parrot gets the job
- Evaluation
- The teacher-and-student test. Confirm with scores that the student is good enough. → the examiner's scorecard