Qwen3-30B-A3B-alpaca-th-52k-dolly-th-15k-wangchan-instruct-seed-4201

This model is a fine-tuned version of Qwen/Qwen3-30B-A3B on the alpaca-th-52k, the dolly-th-15k and the wangchan-instruct datasets. It achieves the following results on the evaluation set:

Loss: 0.6610

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: 0.0002
train_batch_size: 2
eval_batch_size: 2
seed: 4201
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 results

Training Loss	Epoch	Step	Validation Loss
0.9639	0.1149	10	0.9998
0.8143	0.2299	20	0.8182
0.7525	0.3448	30	0.7768
0.7547	0.4598	40	0.7526
0.7413	0.5747	50	0.7367
0.7274	0.6897	60	0.7239
0.7113	0.8046	70	0.7137
0.7288	0.9195	80	0.7052
0.7122	1.0345	90	0.6981
0.7061	1.1494	100	0.6923
0.6744	1.2644	110	0.6858
0.6716	1.3793	120	0.6792
0.6817	1.4943	130	0.6752
0.6445	1.6092	140	0.6720
0.6698	1.7241	150	0.6689
0.6495	1.8391	160	0.6666
0.6585	1.9540	170	0.6647
0.6069	2.0690	180	0.6641
0.624	2.1839	190	0.6631
0.6531	2.2989	200	0.6625
0.6343	2.4138	210	0.6620
0.6546	2.5287	220	0.6616
0.6153	2.6437	230	0.6613
0.6418	2.7586	240	0.6611
0.6418	2.8736	250	0.6611
0.6329	2.9885	260	0.6610