FAQ

Common questions about the current MoroLM research project and its limitations.

Work in progress: this page describes MoroLM-56M v0.6 and will change as the project moves.

Is MoroLM a wrapper around another AI API?#

No. The project includes its own tokenizer, model architecture, training pipeline and checkpoints. The target runtime is local through llama.cpp.

Can v0.6 chat naturally yet?#

Not reliably. It is a raw base model. It has learned language structure but has not been trained to interpret prompts as instructions or conversations.

Why is the model only 56M parameters?#

The 56M model is deliberately small enough to iterate on a consumer GPU while proving data, tokenizer, training, evaluation and export correctness before spending resources on larger models.

Did v0.1–v0.5 fail because the architecture was wrong?#

The core architecture audit did not reveal a fundamental architecture failure. The decisive bug was in the base dataset labels: they were shifted once in the dataset and again in the model loss, training a two-token-ahead task.

Does v0.6 understand Darija?#

It demonstrates learned Darija structure in both Arabic and Latin scripts and can recombine Moroccan article patterns. That is not the same as robust semantic understanding. Natural held-out Darija evaluation is still required.

Can it be commercialized today?#

No production/commercial readiness claim is made for v0.6. Besides model quality, future commercial release requires final provenance review, cleanup of questionable synthetic influences, safety evaluation and product hardening.

Why not just train a 1B model now?#

A larger model would magnify mistakes and cost. The project intentionally proves the full pipeline on 56M first. Scaling becomes rational only after the data and evaluation system are trustworthy.

Will the docs move to a subdomain?#

Yes. This build intentionally keeps documentation under /docs/ for now. The folder can later be moved to a dedicated documentation subdomain.