Track Lineup: 0:00 "Decoding Mistral AI's Large Language Models" - Devendra Chaplot / Mistral 12:50 "No more bad outputs with structured generation" - Remi Louf / .txt (Outlines) 29:50 "Building SOTA Open Weights Tool Use: The Command R Family" - Sandra Kublik / Cohere 2:37:20 "Training Albatross: An Expert Finance LLM" - Leo Pekelis / Gradient 2:56:39 "Fixing bugs in Gemma, Llama & Phi-3" - Daniel Han / Unsloth 3:15:16 "Everything you need to know about Finetuning and Merging LLMs" - Maxime Labonne / Liquid AI
2:41:00 why domain specialized model Using in-context learning to address hallucination is the best method. Therefore we need longer context length. Enterprises need about 1000 few shots to be reliable.
3:01:00 why phi-3 model does show an infinite generations: It's because EOS token and PAD token were the same. Dam... You nailed it. This is so useful. I was so annoyed by Phi-3 for this particular issue.