opt-babylm2-clean-20-epochs-earlystop_seed-42_1e-3
This model was trained from scratch on the kanishka/babylm2-clean dataset. It achieves the following results on the evaluation set:
- Loss: 2.7611
- Accuracy: 0.4673
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.001
- train_batch_size: 32
- eval_batch_size: 64
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 256
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 32000
- num_epochs: 20.0
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Accuracy |
---|---|---|---|---|
4.24 | 0.9998 | 2198 | 3.9528 | 0.3447 |
3.5581 | 2.0 | 4397 | 3.4120 | 0.3948 |
3.2191 | 2.9998 | 6595 | 3.1806 | 0.4180 |
3.0495 | 4.0 | 8794 | 3.0720 | 0.4287 |
2.9458 | 4.9998 | 10992 | 3.0054 | 0.4354 |
2.8669 | 6.0 | 13191 | 2.9663 | 0.4399 |
2.8256 | 6.9998 | 15389 | 2.9382 | 0.4430 |
2.79 | 8.0 | 17588 | 2.9199 | 0.4452 |
2.7624 | 8.9998 | 19786 | 2.9052 | 0.4468 |
2.7361 | 10.0 | 21985 | 2.8915 | 0.4482 |
2.7354 | 10.9998 | 24183 | 2.8843 | 0.4491 |
2.7225 | 12.0 | 26382 | 2.8777 | 0.4500 |
2.7092 | 12.9998 | 28580 | 2.8708 | 0.4505 |
2.6987 | 14.0 | 30779 | 2.8688 | 0.4509 |
2.6894 | 14.9998 | 32977 | 2.8542 | 0.4527 |
2.6561 | 16.0 | 35176 | 2.8258 | 0.4564 |
2.6055 | 16.9998 | 37374 | 2.8005 | 0.4595 |
2.5464 | 18.0 | 39573 | 2.7814 | 0.4627 |
2.4778 | 18.9998 | 41771 | 2.7630 | 0.4658 |
2.4036 | 19.9955 | 43960 | 2.7611 | 0.4673 |
Framework versions
- Transformers 4.45.1
- Pytorch 2.4.1+cu121
- Datasets 3.0.1
- Tokenizers 0.20.0
- Downloads last month
- 14
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.