REFLECTION 70B: What Happened? (#1 on HuggingFace)
Reflection 70B is currently the number #1 trending AI model on HuggingFace, but there is a sequence of events to understand and most importantly, that there is a human on the other side.
Behind every new idea and every new AI model are human beings. And I do admire those persons, trying to provide new open-source models to the AI community. But humans, like I, make mistakes. So it takes time to improve. Without mistakes we would be limited to our current self ...
I am looking forward to examine the new REFLECTION Lllama 3.1 70B model by Matt in some weeks time. Best wishes to Matt to recover soon.
In the second part I explain the simple concept of Strategic Chain-of-Thoughts (SCoT) for improved causal reasoning as published recently and the performance improvements (benchmark data) you can expect, given your specific knowledge domain and your complexity level for causal reasoning on your deployed LLM.
The task to show you the performance of Strategic CoT is a simple math question ”compute the sum of all integers s in the interval −26 to 24”, and we examine both solutions, the one given by standard Chain-of-Thoughts and the other one by the new Strategic CoT.
You understand the at the main idea of SCoT is not "amazing", because we humans perform better when we have a solid knowledge and some strategies about how to handle a topic, but for our little AI minions we have simply to help them in their causal reasoning. And especially for smaller LLM (SLM) this seems an alternative worth to explore.
All rights with authors:
Strategic Chain-of-Thought: Guiding Accurate Reasoning in LLMs through
Strategy Elicitation
https://arxiv.org/pdf/2409.03271
00:00 Top trending AI model: REFLECTION 70B
00:55 Reflection-Tuning LLama 3.1 explained
02:13 Independent Evaluation Reflection 70B
03:33 Ollama weights not correct for Reflection 70B
04:17 HF weights have issues - Retraining Reflection 70B
06:22 Personal statement by Family
08:00 Strategic Chain-of-Thoughts NEW PAPER
08:45 ChatGPT-4o experiment w/ SCoT
13:12 Three examples of SCoT
13:52 Benchmark and Results SCoT
14:34 Free templates for Strategic CoT for YOU
#airesearch
#ai
#insights
Смотрите видео REFLECTION 70B: What Happened? (#1 on HuggingFace) онлайн без регистрации, длительностью часов минут секунд в хорошем качестве. Это видео добавил пользователь Discover AI 08 Сентябрь 2024, не забудьте поделиться им ссылкой с друзьями и знакомыми, на нашем сайте его посмотрели 5,515 раз и оно понравилось 157 людям.