Videos / AI

Episode 8 of 22 · AI, end to end

Hugging Face end to end: models, datasets, spaces

Chapters

1. Three kinds, one thing0:00
2. Clone it and look0:19
3. The pointer and the bytes0:36
4. The card's front matter0:53
5. Pin the revision1:10
6. The cache does the work1:30
7. Open the box1:48
8. Config picks the class2:07
9. Pickles and safetensors2:26
10. The tokenizer is not spare2:44
11. The card is the doc3:03
12. The licence decides3:19
13. A table, not a folder3:38
14. Splits and slices3:57
15. Arrow, memory mapped4:15
16. Stream the river4:33
17. Map, batched4:50
18. The rest of the verbs5:09
19. One socket, every plug5:28
20. What from_pretrained does5:47
21. Text becomes integers6:04
22. Trunk and heads6:22
23. Generate is a loop6:39
24. What pipeline hides6:57
25. Sanding, not felling7:15
26. The loop you do not write7:32
27. The four you supply7:51
28. Two thin matrices8:09
29. Adapters, swapped8:27
30. More than one device8:43
31. The demo booth9:02
32. Hardware, secrets, sleep9:19
33. An endpoint instead9:37
34. What a score actually is9:56
35. Contamination10:13
36. Read it like a sceptic10:31
37. The whole thing10:49

Hugging Face end to end: models, datasets, spaces. Start with a filing cabinet where nothing is ever thrown away. That is the Hugging Face Hub underneath. Every model, every dataset, every demo is a git repository. Ordinary version

Mark as watchedSaved in this browser only.