Hey, thanks for the video ! 🙂 I'm noticing some words cutting out now and then - I'm sure you're aware too just wanted to mention it ! Again, thanks for everything 🙂
In the section “A meditation mantra is all you need” you got better performance by adjusting the system prompt. I’m not very good at prompt engineering. Do you think that I could get comparable performance improvements if instead of playing with the system prompt, I asked LLM to reflect on or revise its answer? This would be kind of like using an agentic workflow. I’m wondering if using reflection would allow me to avoid prompt engineering for each new model I want to try out 🧐🤨
yes though you want to send in the tool output for the next call then you don't want it using the stop token there. You want it generating to the next tool or decision etc.
Claude models use XML tags for outputting and for partitioning things like exemplars for FewShot Learning etc. I cover this in the video about Claude Haiku with examples.
Hey, thanks for the video ! 🙂 I'm noticing some words cutting out now and then - I'm sure you're aware too just wanted to mention it ! Again, thanks for everything 🙂
In the section “A meditation mantra is all you need” you got better performance by adjusting the system prompt.
I’m not very good at prompt engineering. Do you think that I could get comparable performance improvements if instead of playing with the system prompt, I asked LLM to reflect on or revise its answer? This would be kind of like using an agentic workflow. I’m wondering if using reflection would allow me to avoid prompt engineering for each new model I want to try out 🧐🤨
very interesting, best video of gemma 1.1 ;)
thanks much appreciated
Can‘t you tell to send the „stop-token“ after the tool decision? 15:46
yes though you want to send in the tool output for the next call then you don't want it using the stop token there. You want it generating to the next tool or decision etc.
@@samwitteveenai 🤦♂️, in deed. I just thought about: how to avoid a very verbose reply.
thanks for aharing🎉.
a question about 13:03 . What do you mean by output things with "tags" as what people are doing with Claude3.?
Claude models use XML tags for outputting and for partitioning things like exemplars for FewShot Learning etc. I cover this in the video about Claude Haiku with examples.
9:00 👀
Great review! I'm surprise that you can read Thai language 😳
ขอบคุณครับ - I lived in Bkk a long time ago
Thanks, super interesting... FYI. seems both gemma 7b colab links bring you to 1.0
weird I just checked it now and the top seems to be going to 1.1 fine. try again and let me know if you are still having problems
I‘m not sure if I should watch the video: a title like ‚Gemma 1.1 finally useable‘ would be more inviting.
I wonder if they might have implemented Quiet-STaR
5:00 For a true comparison, shouldn't the temperature be set to 0?
Thanks for the awesome video