Researchers are exploring the use of evolutionary algorithms to merge models from HuggingFace, creating new capabilities like Japanese understanding. This model merging technique is cost-effective and does not require fine-tuning, offering vast opportunities for creating merged models.
🌊#LaVague is now available on @huggingface #Space as a managed application! You can now try LaVague without installing anything, and play around automating web actions from natural language instructions. The best part of this? All the stack is open-source! We used Hugging Face…
The more than 500k models on @huggingface provide vast opportunities for creating merged models. @SakanaAILabs's Evolutionary Model Merge algorithm automatically finds optimal solutions in the space of possible model merges. H/T @hardmaru for comments https://t.co/9mUXgHcVGz
Model merging is a cost-effective way to combine useful skills of #LLMs, usually without the need for fine-tuning. Which of the following merging methods have you tried? 🤓 https://t.co/WReQyJGN8e #generativeai #artificialintelligence #llm
Our Huggingface demo is online! Welcome to try it out! The code and weight are also updated. Demo: https://t.co/ZWJtD2rbBk; Project page: https://t.co/r5iGYsj0fl; Weight: https://t.co/185oDgQDZo; Github: https://t.co/hmefpJdrmx; ArXiv: https://t.co/8kp9iEUGtV https://t.co/5nmQenmOnJ
Really exciting research on using evolutionary algorithms to find new frankenmerges. I assume that many ML researchers might dismiss model merge science as "picking up the scraps" behind pretraining work. But has anyone seen a serious critique? https://t.co/pUfT0sPBKO
One of the most imaginative LLM papers I've read in a while: use evolution to merge models from HuggingFace to unlock new capabilities, such as Japanese understanding. It's a form of sophisticated model surgery that requires much smaller compute than traditional LLM training. By… https://t.co/aOVdy3JzmC