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.

qwen2_2b_lora_expert_counting-102400

This model is a fine-tuned version of Qwen/Qwen2-VL-2B on a custom dataset with Counting data (~100k samples).

It achieves the following results on the evaluation set:

• Loss: 0.1590

Training hyperparameters

The following hyperparameters were used during training:

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

Training results

Training Loss	Epoch	Step	Validation Loss
0.2592	0.125	100	0.2344
0.223	0.25	200	0.2009
0.1856	0.375	300	0.1820
0.1881	0.5	400	0.1677
0.1773	0.625	500	0.1606
0.1772	0.75	600	0.1600
0.1731	0.875	700	0.1592
0.1714	1.0	800	0.1590

Framework versions

• PEFT 0.15.2
• Transformers 4.52.4
• Pytorch 2.7.1+cu126
• Datasets 3.6.0
• Tokenizers 0.21.1