Zeitpunkt Nutzer Delta Tröts TNR Titel Version maxTL Do 25.07.2024 00:00:14 231.200 +105 17.196.883 74,4 Mastodon 🐘 4.3.0... 500 Mi 24.07.2024 00:01:07 231.095 +96 17.181.363 74,3 Mastodon 🐘 4.3.0... 500 Di 23.07.2024 00:01:06 230.999 +112 17.166.873 74,3 Mastodon 🐘 4.3.0... 500 Mo 22.07.2024 00:01:10 230.887 +66 17.148.519 74,3 Mastodon 🐘 4.3.0... 500 So 21.07.2024 00:01:06 230.821 +76 17.133.203 74,2 Mastodon 🐘 4.3.0... 500 Sa 20.07.2024 00:00:04 230.745 +27 17.120.900 74,2 Mastodon 🐘 4.3.0... 500 Fr 19.07.2024 13:57:31 230.718 +146 17.111.706 74,2 Mastodon 🐘 4.3.0... 500 Do 18.07.2024 00:00:59 230.572 +90 17.086.626 74,1 Mastodon 🐘 4.3.0... 500 Mi 17.07.2024 00:01:10 230.482 +55 17.071.683 74,1 Mastodon 🐘 4.3.0... 500 Di 16.07.2024 00:01:08 230.427 0 17.055.997 74,0 Mastodon 🐘 4.3.0... 500
Hobson Lane (@hobs) · 07/2022 · Tröts: 3.977 · Folger: 1.028
Do 25.07.2024 04:29
1. Use something like axolotl to record your hyperparameters and compare them to what others are doing
2. Instrument your testing so you can reuse all that labeled data you are implicitly labeling each time you or your users interact with an #llm.
3. Curate your #evaluation and #training datasets to make sure it matches your real world use cases and what your users are actually doing and you aren't tuning on garbage #data, you can get great results with < 1000 examples
2/2
[Öffentlich] Antw.: 0 Wtrl.: 1 Fav.: 0 · via Fedilab