SmolKalam: Ensemble Quality-Filtered Translation at Scale for High Quality Arabic Post-Training Data
This addresses the need for better Arabic datasets for post-training in language models, though it appears incremental as it builds on existing translation approaches.
The researchers tackled the lack of large-scale, high-quality Arabic datasets for post-training by introducing SmolKalam, a translation of Smoltalk2 that uses an ensemble translation pipeline with quality filtering, achieving improved dataset quality through systematic ablations.
Although the community has tackled the acquisition of high-quality Arabic pretraining data, we still lack large-scale, multi-turn Arabic datasets that include reasoning and tool calling. Naive translation can work at the pretraining scale, but post-training demands much higher quality, which requires a stricter approach to dataset curation. In this work, we introduce SmolKalam, a translation of Smoltalk2 that uses a multi-model ensemble translation pipeline, applies quality filtering, and examines effective translation techniques for traditional decoder-only models through ablations.