This is an encoder Language Model pre-trained from scratch on transcriptions of the archives of the Dutch East India Company. It is therefore a model specialized on Early Modern Dutch as used in the archive (1602–1800). The model follows a RoBERTa architecture. It can be fine-tuned on any NLP task.
This version of the model is the best performing GloBERTise model when tested on binary event detection of the four I have pre-trained (in august 2025)
Comparison to other models: Adapted settings for 'num_training_steps' and 'num_warmup_steps' compared to GloBERTise-v01 and GloBERTise-v01-rerun, otherwise the same. Different seed compared to GloBERTise-rerun, same parameter settings.
See my GitHub repos
- for pre-training: https://github.com/globalise-huygens/GloBERTise
- for evaluation: https://github.com/globalise-huygens/GloBERTise-eval
And a small presentation: https://docs.google.com/presentation/d/1gkg5hChWAMXA6mxfgFkkvIieWdj_17yKitwBkBNcJBo/edit?usp=sharing
Most important parameter settings:
| learning rate | 0.0003 |
| betas | [ 0.9, 0.98] |
| weight_decay | 0.01 |
| num_train_epochs | 2 |
| per_device_train_batch_size | 40 |
| gradient_accumulation_steps | 10 |
| fp16 | true |
I pre-trained this model on National Supercomputer Snellius, which we were granted access to by the Dutch national e-infrastructure with the support of the SURF Cooperative using grant no. EINF-11854
I calculated a rough estimate of how much carbon was emitted by pre-training this model with https://mlco2.github.io/impact/#publish. Experiments were conducted using a private infrastructure, which has a carbon efficiency of 0.432 kg CO2/kWh. A cumulative of 24 hours of computation was performed on hardware of type A100 PCIe 40/80GB (TDP of 250W). Total emissions are estimated to be 2.59 kg CO2 of which 0 percents were directly offset.