Tag
This paper tackles code generation for no-resource programming languages by building benchmarks and proposing a method that combines further pre-training with weight difference transfer to create specialized instruction-following models at reduced cost.