llava-1.5-7b-hf-meme-lora

<style> .img-responsive { width: 100%; height: auto; } </style>

Conclusion

While significantly better at understanding and describing emotions and details in images compared to LLaVA-1.5-7b-hf, the fine-tuned model struggles with recognizing text.

Train Loss

<img src="./adapter-module/training_loss.png" alt="loss" class="img-responsive">

Test

A comparative analysis of emoji in prompts, differents between the original model and its fine-tuned counterpart. </br> Original Model:https://huggingface.co/llava-hf/llava-1.5-7b-hf/</br> <img src="./images/original-01.JPG" alt="meme01" class="img-responsive"> <img src="./images/original-02.JPG" alt="meme02" class="img-responsive"> <img src="./images/original-03.JPG" alt="meme03" class="img-responsive">

Fine-tuned Lora Model:https://huggingface.co/REILX/llava-1.5-7b-hf-meme-lora</br> <img src="./images/lora-01.JPG" alt="meme01" class="img-responsive"> <img src="./images/lora-02.JPG" alt="meme02" class="img-responsive"> <img src="./images/lora-03.JPG" alt="meme03" class="img-responsive">

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• cutoff_len: 2048
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 8
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 5.0