Observed arrival · 2026-10-17
MoroLM: A Darija model that shows its work
A from-scratch Moroccan Darija language-model research project, with public documentation of its training pipeline, current limits, and development history.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Researchers and developers studying Moroccan Darija language models
- Worth noticing
- A label-shift bug made earlier checkpoints predict two tokens ahead; v0.6 restarted from random weights.
Field notes
The documentation breaks the work into source audit, cleaning, tokenization, pretraining, evaluation, and export, with each stage marked by its current status. It reports that the v0.6 run used about one pass over a roughly 101-million-token corpus and cautions that its reported perplexity of 185 matches generations that remain far from fluent. The page also distinguishes this clean base-model lineage from earlier instruction-tuning probes.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue