This paper presents R-BPE, a lightweight framework for adapting existing Byte-Pair Encoding (BPE) tokenizers to better support a specified target language.It reuses tokens from user-excluded languages and creates ID-based maps to resolve the new tokens of the chosen language.We evaluate R-BPE on Arabic as a target language. R-BPE reduced subword fertility by an average of 24.4% across the LLaMA 3.1 8B, Command R 35B, and Qwen 3 8B models.Applied to LLaMA 3.1 8B in continued pretraining mode, R-BPE yields a 7.33% reduction in training time. On the ArabicMMLU benchmark, the resulting model improved by 5.09 points on five in-domain topics and matched the original model’s overall performance.It also preserved performance on EnglishMMLU. R-BPE effectively leverages existing models’ tokenizers, embedding layers, and performance to better support target languages without incurring model size changes. We release an R-BPE implementation that is compatible with HuggingFace interfaces and thereby readily applicable to a wide range of existing models at https://acr.ps/1L9GPmL.

In this paper, we present our submission for the WojoodNER 2024 Shared Tasks addressing flat and nested sub-tasks (1, 2). We experiment with three different approaches. We train (i) an Arabic fine-tuned version of BLOOMZ-7b-mt, GEMMA-7b, and AraBERTv2 on multi-label token classifications task; (ii) two AraBERTv2 models, on main types and sub-types respectively; and (iii) one model for main types and four for the four sub-types. Based on the Wojood NER 2024 test set results, the three fine-tuned models performed similarly with AraBERTv2 favored (F1: Flat=.8780 Nested=.9040). The five model approach performed slightly better (F1: Flat=.8782 Nested=.9043).

This paper details our submission to the WojoodNER Shared Task 2024, leveraging in-context learning with large language models for Arabic Named Entity Recognition. We utilized the Command R model, to perform fine-grained NER on the Wojood-Fine corpus. Our primary approach achieved an F1 score of 0.737 and a recall of 0.756. Post-processing the generated predictions to correct format inconsistencies resulted in an increased recall of 0.759, and a similar F1 score of 0.735. A multi-level prompting method and aggregation of outputs resulted in a lower F1 score of 0.637. Our results demonstrate the potential of ICL for Arabic NER while highlighting challenges related to LLM output consistency.