Fine-tuning Gemma 3-4B and 27B on about 50,000 mixed Darija and English instructions produces a 27B model that matches Atlas-Chat on DarijaMMLU and exceeds it on DarijaHellaSwag, using 48 GPU-hours.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data
Fine-tuning Gemma 3-4B and 27B on about 50,000 mixed Darija and English instructions produces a 27B model that matches Atlas-Chat on DarijaMMLU and exceeds it on DarijaHellaSwag, using 48 GPU-hours.