Videos / AI

Episode 18 of 22 · AI, end to end

Fine-tuning a model for one real use case

Chapters

1. One note, four fields0:00
2. Nine notes in ten0:17
3. Almost right is costly0:34
4. Why this task fits0:52
5. Three ways to not1:12
6. Describe or demonstrate1:29
7. The labels already exist1:46
8. A label you checked2:03
9. Deduplicate first2:20
10. Split it honestly2:40
11. The golden set, frozen2:56
12. The shape of one example3:14
13. Six dials3:32
14. Pick the base model3:48
15. What an adapter is4:06
16. Rank, and alpha4:24
17. Learning rate and epochs4:43
18. Batch, and accumulation5:02
19. Two lines on a chart5:21
20. A healthy run5:38
21. The elbow, and after5:55
22. Two other bad shapes6:13
23. Checkpoints, and choosing6:30
24. Read the samples too6:48
25. A blind tasting7:06
26. Four numbers7:22
27. Give the base a fair fight7:40
28. The leaderboard cannot help7:57
29. Read the failures8:16
30. The bar, written first8:32
31. The old part stays in the van8:49
32. Merge, or keep it separate9:04
33. Quantise, then retest9:21
34. Version weights with data9:39
35. Shadow, then canary9:57
36. It decays10:14
37. The whole build10:32

Fine-tuning a model for one real use case. A lift engineer finishes a job and types a note on her phone. Somewhere downstream that note has to become four boxes on a parts order: part code, quantity, urgency, site. Four box

Mark as watchedSaved in this browser only.