MoroLMRoadmap
The next technical milestones from v0.6 toward useful Darija applications.
Work in progress: this page describes MoroLM-56M v0.6 and will change as the project moves.
Immediate research sequence#
- Natural Darija holdout: build evaluation prompts that are not locality/article templates.
- Data audit: reduce repeated template families and clean suspicious synthetic/replay concentrations.
- Base checkpoint decision: decide whether v0.6 receives more clean pretraining or freezes as the SFT parent.
- Instruction tuning: create a high-quality Darija-first assistant dataset with proper system/user/assistant formatting.
- Evaluation harness: separate fluency, relevance, factuality, code-switching and instruction-following scores.
- Export: produce verified GGUF builds and benchmark local latency.
Scale path#
The 56M model is a pipeline proof. The longer-term ambition is to move toward substantially larger models — potentially around 1B–2B parameters — only after data quality, training correctness, evaluation and deployment are stable.
Potential applications#
Education#
Private, on-premise university assistants that search course material and help with small student tasks.
Moroccan services#
Darija interfaces for local institutions and businesses where language fit matters.
Local privacy#
Deployments where prompts, documents and student/customer data remain on the institution's own infrastructure.
Developer platform#
A compact Moroccan language layer that can later expose local APIs and browser-based tooling.