daviBera/intern35_2b_lora_expert_counting-102400

<h2 align="center"> Linear Model Merging Unlocks Simple and Scalable Multimodal Data Mixture Optimization

<br>

   [](https://github.com/BerasiDavide/mLLMsmerging4DMO) </h2>

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.

intern352bloraexpertcounting-102400

This model is a fine-tuned version of OpenGVLab/InternVL3_5-2B-Pretrained-HF on a custom dataset with Counting data (~100k samples).

It achieves the following results on the evaluation set:

• —Loss: 0.2169

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 2e-05
• —trainbatchsize: 8
• —evalbatchsize: 4
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 4
• —gradientaccumulationsteps: 4
• —totaltrainbatch_size: 128
• —totalevalbatch_size: 16
• —optimizer: Use adamwtorch with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_ratio: 0.1
• —training_steps: 800

Training results

Framework versions

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