halcyon-llm/SmolLM2-360M-japanese_base_phase_2-11403

scratch_adamw_phase_2

This model is a fine-tuned version of halcyon-llm/SmolLM2-360M-japanese_base_phase_1-11000 on the kajuma/training_01-09_token dataset. It achieves the following results on the evaluation set:

• Loss: 1.1349

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: 0.003
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 64
• total_train_batch_size: 256
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine_with_min_lr
• lr_scheduler_warmup_steps: 1000
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.1622	0.0438	500	1.1753
1.2027	0.0877	1000	1.2353
1.207	0.1315	1500	1.2419
1.2536	0.1754	2000	1.2374
1.2442	0.2192	2500	1.2315
1.2163	0.2631	3000	1.2262
1.1969	0.3069	3500	1.2185
1.2102	0.3508	4000	1.2099
1.1908	0.3946	4500	1.2016
1.1611	0.4385	5000	1.1929
1.2055	0.4823	5500	1.1835
1.1743	0.5262	6000	1.1750
1.1519	0.5700	6500	1.1674
1.1397	0.6138	7000	1.1601
1.1702	0.6577	7500	1.1538
1.1633	0.7015	8000	1.1483
1.1248	0.7454	8500	1.1439
1.1306	0.7892	9000	1.1407
1.0726	0.8331	9500	1.1385
1.1104	0.8769	10000	1.1369
1.0933	0.9208	10500	1.1359
1.0918	0.9646	11000	1.1352

Framework versions

• Transformers 4.48.0.dev0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0