nmatheg v 2.0

- [ ] fix tokenizer save vs. save_model
- [ ] merge the logic of finetune and pretrain
```
{'WordTokenizer': {'10000': {'off-eval-ar': {'birnn': {'loss': [0.6514959335327148],
'precision': [0.5062719851330723],
'recall': [0.5004618023485948],
'f1': [0.454612213786187],
'accuracy': [0.7950155763239876]}}}}}
```
vs.

```
{'UBC-NLP/MARBERTv2': {'off-eval-ar': {'loss': [0.4560829699039459],
'precision': [0.4009345794392523],
'recall': [0.5],
'f1': [0.44502074688796683],
'accuracy': [0.8018691588785046]}}}
```

- [ ] prevent data leakage between different runs/models
- [ ] test multiple tasks

Start by locating the tokenizer save and save_model paths, then trace the separate finetune and pretrain entry points. Review how runs and models share data, and identify the existing task-evaluation flow. Done means the logic is unified, runs do not leak data, and multiple tasks are covered by tests or evaluations.

Written by the indexing model from the issue text.