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).

qwen2_2b_lora_expert_chart-102400

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

It achieves the following results on the evaluation set:

• Loss: 0.9304

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
1.0359	0.125	100	0.9610
0.9983	0.25	200	0.9451
1.0097	0.375	300	0.9390
1.0007	0.5	400	0.9355
1.0301	0.625	500	0.9341
0.9492	0.75	600	0.9316
0.9425	0.875	700	0.9305
0.989	1.0	800	0.9304

Framework versions

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