Adapting large language models to specialized domains remains challenging due to the computational cost of full finetuning and the limited availability of domain-specific parallel data. We present a systematic framework for parameter-efficient domain adaptation using Low-Rank Adaptation (LoRA) geared towards efficient learning in low-resource scenarios. Our method combines data-scaling analysis, dual-track hyperparameter optimization, and competitive benchmarking. We evaluate our approach on the low-resource English–Norwegian petroleum translation domain using a distilled version of NLLB and parallel data from the Norwegian Petroleum Directorate. Our adapted model achieves 61.48 BLEU (+24.62 over the base model) and 0.9298 COMET, while updating <0.4% of parameters. Our results provide a reproducible and computationally efficient blueprint for domain adaptation in neural machine translation, particularly for specialized and resource-constrained domains.

A graph is called supereulerian if it has a spanning eulerian subgraph. A graph is said to be hamiltonian if it has a spanning cycle. A nontrivial path is called a branch if it has only internal vertices of degree two and end vertices of degree not two. Let S be a set of branches of G, then S is called a branch cut if G−S has more components than G. A minimal branch cut is called a branch-bond. In this paper, we characterize one or pairs of those forbidden subgraphs that force a 2-edge-connected graph satisfying that every odd branch-bond has an edge branch to be supereulerian. We also characterize one or pairs of those forbidden subgraphs that force a 2-connected supereulerian graph to be hamiltonian.

Map, Explore & Exploit (ME2) is a scalable meta-heuristic for problems in the field of multi-modal, multi-dimension optimization. It has a modular design with three phases, as reflected by its name. Its first phase (Map) generates a set of samples that is mostly uniformly distributed over the search space. The second phase (Explore) explores the neighbourhood of each sample point using an evolutionary strategy, to find a good - not necessarily optimal - set of neighbours. The third phase (Exploit) optimizes the results of the second phase. This final phase applies a simple gradient descent algorithm to find the local optima for each and all of the neighbourhoods, with the objective of finding a/the global optima of the whole space. The performance of ME2 is compared, on a fair basis, with the performance of benchmark optimization algorithms: Genetic Algorithms, Particle Swarm Optimization, Simulated Annealing and Covariance Matrix Adaptation Evolution Strategy. In most test cases it finds the global optima earlier than the other algorithms. It also scales-up, without loss of performance, to higher dimensions. Due to the distributed nature of ME2’s second and third phase, it can be comprehensively parallelized. The search & optimization process during these two phases can be applied to each sample point independently of all the others. A multi-threaded version of ME2 was written and compared to its single-threaded version, resulting in a near-linear speed-up as a function of the number of cores employed.