MoroLMModel card
Model card for MoroLM-56M Base v0.6 Cleanstart.
Work in progress: this page describes MoroLM-56M v0.6 and will change as the project moves.
Model summary#
| Field | Value |
|---|---|
| Name | MoroLM-56M Base v0.6 Cleanstart |
| Type | Decoder-only causal language model |
| Parameters | 56,607,232 |
| Context configured | 4,096 positions |
| Training block length | 512 |
| Primary target | Moroccan Darija |
| Additional languages | English, French, Arabic/MSA, code-switching |
| Status | Experimental base model |
Intended use today#
Research, pipeline validation, tokenizer/model experiments, local inference experiments and future instruction-tuning work.
Not intended today#
- Production customer support or institutional deployment.
- High-stakes factual, legal, medical or financial advice.
- Reliable retrieval of facts about Morocco.
- Chat use without instruction tuning and guardrails.
Known limitations#
- Strong repetition under greedy decoding.
- Hallucinated entities and incorrect facts.
- Topic drift from short prompts.
- Training-template memorization and recombination.
- Uneven strength between Arabic-script Darija and Latin Darija.
- Only ~102M token-position exposures for a 56.6M-parameter multilingual model.
Do not market v0.6 as a reliable assistant. It is the first clean base milestone, not the finished product.