Four levels of adaptation of open-source models

Medium about the difficulties and problems of LLM.

How to determine the required level of adaptation

To better understand the differences between the approaches, we calculated the time it takes to create a Proof-of-Concept (PoC). Each level requires its own amount of resources:

The most difficult and time-consuming stage is the last one. It is easier not to bother with additional training, but to look for a faster and simpler approach. But it is not a fact that this solution will be the best. If you are stuck at one stage longer than it takes to test the hypothesis, it is worth admitting: the current method does not work. You need to complicate the solution or combine techniques from different approaches – for example, training and agents. In any case, do not get hung up on one level and do not try to squeeze the impossible out of it.

If you are faced with adapting models, remember:

one way or another, both cloud and open source models will have to be adapted;
Some of the problems that arise can be solved by prompt engineering. The main thing is not to overdo it and stop in time;
With the help of agents you can add new features that are not available by default;
retraining is a complex and expensive process. It is necessary if you cannot cope with the task at the previous levels of adaptation;
The best solution for specific problems is not to stick to one approach, but to combine different ones.

That's all I have for now, but I'm ready to answer your questions. Thanks for reading!