Beetle publishes 330 small models across 21 languages, thousands of checkpoints, training data and curriculum-learning code. It compares five curricula at three token budgets and reports that staged curricula better match human second-language reading times.
Beetle publishes a multilingual suite for testing learning curricula
2026-10-24Verified announcementTRIAL
Beetle publishes 330 small models across 21 languages, thousands of checkpoints, training data and curriculum-learning code. It compares five curricula at three token budgets and reports that staged curricula better match human second-language reading times.
What this means for you
The artifacts turn multilingual training-order claims into reproducible experiments. Results come from small models and controlled corpora and should not be extrapolated directly to frontier-scale systems.
Evidence & related material
Technical details & scores
| Artifacts | 330 models, thousands of checkpoints, data and code |
|---|---|
| Languages | 21 |
| Curricula | 5 |
| Training Budgets | 100M, 2B and 24B tokens |
| Publication | EMNLP 2026 oral |
| Limitation | Small-model controlled study; transfer to frontier-scale training is unverified |
| Time Basis | Official pages provide only 2026-10-24 date precision; ACL build metadata shows public availability by 10:04Z, before the overlap start. |
Technical impact 3
Engineering value 3
Adoption signal 1
Maturity 2
Verification cost 2
Risk 1