LibraCaption/InternVL3-38B-Captioner

sft_captioner

This model is a fine-tuned version of OpenGVLab/InternVL3-38B-hf on the pyq_part1_captioner_0815, the pyq_part2_captioner_0815 and the private_captioner_0815_optimized.json datasets. It achieves the following results on the evaluation set:

• Loss: 0.7323

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: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 2025
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 32
• total_eval_batch_size: 32
• optimizer: Use 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: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.8966	0.0868	500	0.8954
0.861	0.1735	1000	0.8705
0.8622	0.2603	1500	0.8549
0.82	0.3470	2000	0.8349
0.8211	0.4338	2500	0.8202
0.7978	0.5206	3000	0.8064
0.7955	0.6073	3500	0.7945
0.7845	0.6941	4000	0.7838
0.7617	0.7808	4500	0.7729
0.7772	0.8676	5000	0.7622
0.7641	0.9544	5500	0.7544
0.6061	1.0411	6000	0.7635
0.5863	1.1279	6500	0.7613
0.5777	1.2146	7000	0.7588
0.5943	1.3014	7500	0.7490
0.5816	1.3882	8000	0.7469
0.5723	1.4749	8500	0.7421
0.5721	1.5617	9000	0.7374
0.5724	1.6484	9500	0.7353
0.5731	1.7352	10000	0.7343
0.5597	1.8220	10500	0.7330
0.5731	1.9087	11000	0.7326
0.5557	1.9955	11500	0.7323

Framework versions

• Transformers 4.52.4
• Pytorch 2.7.1+cu126
• Datasets 3.6.0
• Tokenizers 0.21.1