This paper came out recently: https://arxiv.org/abs/2207.14502 . It shows a way to work around the lack of sufficient training data for generating computer programs by “generating synthetic programming puzzles and solutions, verified for correctness by a Python interpreter.” We can think of analogous generation for data-limited general LLMs and there are some possibilities.
This paper came out recently: https://arxiv.org/abs/2207.14502 . It shows a way to work around the lack of sufficient training data for generating computer programs by “generating synthetic programming puzzles and solutions, verified for correctness by a Python interpreter.” We can think of analogous generation for data-limited general LLMs and there are some possibilities.