learning_rate: 0.0002
train_batch_size: 2
eval_batch_size: 2
seed: 42
distributed_type: multi-GPU
num_devices: 64
gradient_accumulation_steps: 8
total_train_batch_size: 1024
total_eval_batch_size: 128
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: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 3.0

Training Loss	Epoch	Step	Validation Loss
0.7193	0.2326	10	0.2340
0.0363	0.4651	20	0.0410
0.0272	0.6977	30	0.0311
0.0287	0.9302	40	0.0277
0.018	1.1628	50	0.0271
0.0191	1.3953	60	0.0257
0.0162	1.6279	70	0.0257
0.0191	1.8605	80	0.0256
0.0113	2.0930	90	0.0255
0.013	2.3256	100	0.0270
0.0096	2.5581	110	0.0274
0.0091	2.7907	120	0.0275

airesearch/Qwen3-30B-A3B-medqa

Qwen3-30B-A3B-medqa