Linear Model Merging Unlocks Simple and Scalable Multimodal Data Mixture Optimization

This is an official checkpoint from the paper: "Linear Model Merging Unlocks Simple and Scalable Multimodal Data Mixture Optimization " (link). See the official implementation for more information on how to use the models.

intern35_8b_lora_expert_counting-102400

This model is a fine-tuned version of OpenGVLab/InternVL3_5-8B-Pretrained-HF on a custom dataset with Counting data (~100k samples).

It achieves the following results on the evaluation set:

Loss: 0.1607

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 8
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 4
total_train_batch_size: 128
total_eval_batch_size: 4
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
training_steps: 800

Training results

Training Loss	Epoch	Step	Validation Loss
0.2191	0.125	100	0.2157
0.2017	0.25	200	0.1917
0.161	0.375	300	0.1735
0.1797	0.5	400	0.1705
0.1715	0.625	500	0.1656
0.1719	0.75	600	0.1636
0.176	0.875	700	0.1613
0.1713	1.0	800	0.1607