qwen3-4B-instruct-refiner-sft

This model is a fine-tuned version of Qwen/Qwen3-4B-Instruct-2507 on the refiner_sft_hard_filtered_train dataset. It achieves the following results on the evaluation set:

• Loss: 1.1232

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: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• 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: 0.05
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss
0.4937	0.1874	100	0.6320
0.511	0.3749	200	0.6321
0.4657	0.5623	300	0.6459
0.4577	0.7498	400	0.6420
0.4634	0.9372	500	0.6470
0.2661	1.1256	600	0.6921
0.2427	1.3130	700	0.6904
0.2608	1.5005	800	0.6896
0.2811	1.6879	900	0.6763
0.2506	1.8754	1000	0.6782
0.1031	2.0619	1100	0.7820
0.1053	2.2493	1200	0.7939
0.1009	2.4367	1300	0.7773
0.1022	2.6242	1400	0.7983
0.1087	2.8116	1500	0.8067
0.1046	2.9991	1600	0.8037
0.0311	3.1856	1700	0.9448
0.0343	3.3730	1800	0.9443
0.0322	3.5604	1900	0.9526
0.0299	3.7479	2000	0.9680
0.0335	3.9353	2100	0.9606
0.0073	4.1218	2200	1.0976
0.0069	4.3093	2300	1.1145
0.0064	4.4967	2400	1.1218
0.0086	4.6842	2500	1.1228
0.0072	4.8716	2600	1.1233

Framework versions

• Transformers 4.52.4
• Pytorch 2.10.0+cu128
• Datasets 4.8.4
• Tokenizers 0.21.1