The situation with artificial intelligence is the order of the day, we can’t stop talking about it and everyone is fascinated by what this technology can achieve. Microsoft, Google, Meta, OpenAI are some of the great companies behind what we enjoy every day. Although open source artificial intelligence, they are putting up a good fight and are an alternative to consider. Now Goalwho seeks to face ChatGPT and Bardplan to do it with LIMEa powerful AI which would be a variant of the LLaMa model.
Actually, LIMA is quite an interesting variant of Meta’s LLaMa model. As we can see in a study For meta, the language models are trained in two stages. The first of these stages consists of unsupervised training that uses a fair amount of text. Without any kind of filter that allows learning within a more general scenario.
While the second stage is a kind of debugging of the previous one. Where it is trained in more detail for certain specific tasks or user preferences.
LIMA: a language model based on LLaMa
So Meta did just that with LIMA, an AI looking to take on OpenAI’s ChatGPT and Google’s Bard. Those who are making millionaire investments in this technology to stay at the top of the sector. Obviously, Meta does not want to, nor can he be left behind.
As we discussed earlier, Less Is More for Alignment (LIMA) It is a language model that is based on LLaMa with 65 million parameters and that with less than 1,000 prompts it manages to behave correctly. The most interesting thing about the matter is that we are talking about the fact that this “reinforcement” or modeling based on human preferences was not necessary and the result was impressive.
This model was developed by Meta, though with significant help from Carnegie Mellon University, the University of Southern California, and Tel Aviv University. Everyone agrees that LIMA’s performance is impressive and that he has managed to follow answer formats with very few examples available for training. It is even possible to perform some tasks for which you were not even trained.
The most interesting thing about this issue is that it reaches the same level as ChatGPT or Bard, it could even be said that it surpasses them. Taking into account that LIMA’s responses have been 43% better in the case of GPT-4 and 58% in the case of Bard.
Did we forget about RLHF?
But the situation here takes on larger proportions when, thanks to this study, we realize that the use of the technique “Reinforcement Learning from Human Feedback” (RLHF) is not as essential as we thought.
In a nutshell, this system uses humans to optimize the AI’s results while training. This is an overly expensive process that companies use to make their language models more accurate.
Although, from Meta they assure that in this stage of training it is better “more quality than quantity”. They also comment that developing this series of high-quality data is not exactly easy and is not an option in all cases.