babylm-anon/rta

babylm-base5M-gpt2 (Fork with chck_100M Checkpoint)

This repository is a fork of alexandertam/babylm-base5m-gpt2, created for the BabyLM Challenge 2025 submission. It extends the original model by including an additional checkpoint (chck_100M), adhering to the challenge's guidelines.

Model Description

This model is a pre-trained version of the GPT-2 architecture. It achieves the following results on the evaluation set:

• Loss: 3.0628
• Accuracy: 0.4521

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: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 190
• training_steps: 19000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
5.579	0.1024	200	4.7677	0.3189
4.7716	0.2048	400	4.3385	0.3544
4.5162	0.3072	600	4.1772	0.3593
4.4056	0.4096	800	4.0754	0.3693
4.3138	0.5120	1000	4.0143	0.3626
4.2148	0.6144	1200	3.9601	0.3554
4.1925	0.7168	1400	3.9019	0.3723
4.0293	0.8193	1600	3.8579	0.3749
3.9407	0.9217	1800	3.8101	0.3782
3.8371	1.0241	2000	3.7870	0.3721
3.0659	2.0481	4000	3.4672	0.4085
2.6866	3.0722	6000	3.2850	0.4316
2.5063	4.0963	8000	3.1963	0.4372
2.4139	5.1203	10000	3.1406	0.4442
2.3246	6.1444	12000	3.1152	0.4484
2.3111	7.1685	14000	3.0879	0.4489
2.2761	8.1925	16000	3.0668	0.4542
2.2231	9.2166	18000	3.0695	0.4517

Framework versions

• Transformers 4.50.3
• Pytorch 2.7.1+cu126
• Datasets 3.6.0
• Tokenizers 0.21.4