if you want to test reasoning try this slightly changed riddle: "I hang 7 shirts out to dry in the Sun. After 5 hours all shirts are dry. The next day i hang 14 shirts out to dry. The conditions are the same. How long will it take to dry 14 shirts? take a deep breath and proceed step by step" 99%of LLMs will say it needs 10 hours including gemma-7B. If you change the prompt adding a example riddle (a 1-shot prompt) with a similar structure, the AI can learn the pattern. For example, a riddle about 3 t-shirts drying in 3 hours, then 6 t-shirts drying also in 3 hours, will help the AI understand that 14 t-shirts would only need 5 hours to dry.
>>> i hang 7 shirts out to dry in the Sun. After 5 hours all shirts are dry. The next day i hang 14 shirts out to dry. The conditions are the same. How long will it take to dry 14 shirts? take a deep breath and proceed step by step gemma2b: The total time taken to dry 7 shirts is 5 hours. Since the shirts are hanging in the same conditions, we can assume that the drying process follows the same rate. Therefore, to dry 14 shirts, it will also take 5 hours.
GPT 4 responds correctly to this riddle. "If 7 shirts dry in 5 hours under certain conditions, and the next day the conditions are exactly the same, 14 shirts will also dry in 5 hours, provided they all receive the same exposure to the drying conditions."
Dont forget starcoder and santacoder models. They are among the earliest opensource models that standardized data quality checks and pipelines. And inspired so many new models.
00:00 Introduction of various open source language models 01:19 Google has open-sourced Gemma, a suite of models 02:34 Introducing Gemma - 2B 7B 6Trillion Tokens 03:46 Models trained on TPU V5e with impressive benchmarks. 04:57 Gemma's terms of use and access request process 06:02 Using Keras 3.0 and Keras NLP for NLP models 07:11 Gemma 2B 7B 6 trillion tokens model's potential for multilingual fine-tuning. 08:18 Gemma 2B 7B 6Trillion Tokens for NLP Crafted by Merlin AI.
fyi you say the weights are only English but in my tests it was able to respond to queries in French. It's possible they were going for an English-only dataset but accidentally brought in some other language data.
Yeah this is quite common. Especially with languages like French, Spanish etc. A lot of other languages appear even in english text and when you have 6 Trillion tokens that can add up to. a lot. Also the tokenizer is a multi-lingual tokenizer (like the full size Gemnini models) so this can help as well.
@@samwitteveenai wasnt llama 3 supposed to be really powerful and almost a really really primative "agi" that what i got from that little zuckerburg speech
@@samwitteveenai I don't quite understand llama vs gemma. arent they both models? but why does it sound like gemma would run on top of llama, or how llamacpp allows for any model to be run on it, dont understand the layers here.
@@pylotlight it is just a model (2 different sizes of models) there are versions for cpp and other frameworks so it can run on various frameworks, but at the end of the day both Gemma and LLaMA are models