00:07 Focus on fine-tuning LLMs and model merging. 02:27 Knowing when to use fine-tuning is crucial for model performance and cost. 06:57 Different characteristics of a good dataset 09:13 Improving LLM capabilities through complex tasks 13:48 Training loss behavior and model merging 16:13 Linear interpolation allows precise merging of models 20:41 The speaker shared resources for LLM learning and announced a new book. 23:00 Merging LLMs requires careful selection and understanding of merge techniques. 27:15 Fine-Tuning LLMs can be unpredictable 29:15 Preference alignment helps heal the models for better performance
This was a very clear presentation. How would you suggest doing a demonstration of concept of a digital platform for collective intelligence? The platform would keep humans in the loop through a conversational interface that merges public parts of conversations into a shared graph representation. It would report on which issues are discussed by the most people, aggregate measures of sentiment about proposed solutions, and show how those measures of sentiment change over time. One big challenge is how the platform would authenticate the identities of humans at the same time that it allows people to share selected parts of conversations anonymously.
00:07 Focus on fine-tuning LLMs and model merging.
02:27 Knowing when to use fine-tuning is crucial for model performance and cost.
06:57 Different characteristics of a good dataset
09:13 Improving LLM capabilities through complex tasks
13:48 Training loss behavior and model merging
16:13 Linear interpolation allows precise merging of models
20:41 The speaker shared resources for LLM learning and announced a new book.
23:00 Merging LLMs requires careful selection and understanding of merge techniques.
27:15 Fine-Tuning LLMs can be unpredictable
29:15 Preference alignment helps heal the models for better performance
Where can we download the ppt file?
This was a very clear presentation. How would you suggest doing a demonstration of concept of a digital platform for collective intelligence? The platform would keep humans in the loop through a conversational interface that merges public parts of conversations into a shared graph representation. It would report on which issues are discussed by the most people, aggregate measures of sentiment about proposed solutions, and show how those measures of sentiment change over time. One big challenge is how the platform would authenticate the identities of humans at the same time that it allows people to share selected parts of conversations anonymously.
I'll have to do some DPO now😅