LLaVA-Gemma-2b-LORA

LLaVA-Gemma-2b-LORA is a Vision-Language Model (VLM) based on google/gemma-2b-it model and trained in original LLaVA setup using LORA. This model is primarily adapted to work with Russian, but still capable to work with English.

Usage

Model usage is simple via transformers API

import requests

from PIL import Image
from transformers import AutoProcessor, AutoTokenizer, LlavaForConditionalGeneration

model_name = "deepvk/llava-gemma-2b-lora"

model = LlavaForConditionalGeneration.from_pretrained(model_name)
processor = AutoProcessor.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
img = Image.open(requests.get(url, stream=True).raw)
messages = [
    {"role": "user", "content": "<image>\nОпиши картинку несколькими словами."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(images=[img], text=text, return_tensors="pt")

generate_ids = model.generate(**inputs, max_new_tokens=30)
answer = tokenizer.decode(generate_ids[0, inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(answer)

Use the <image> tag to point to an image in the text and follow the chat template for a multi-turn conversation. The model is capable of chatting without any images or working with multiple images in a conversation, but this behavior has not been tested.

The model format allows it to be directly used in popular frameworks, e.g. you can test the model using lmms-eval, see Results section for details.

Train

To train this model, we follow the original LLaVA pipeline and reuse haotian-liu/LLaVA framework.

The model was trained in two stages:

1. The adapter was trained using pre-training data from ShareGPT4V.
2. Instruction tuning included training the LLM and the adapter, for this we use:
   • deepvk/LLaVA-Instruct-ru — our new dataset of VLM instructions in Russian
   • deepvk/GQA-ru — the training part of the popular GQA test, translated into Russian, we used the post-prompt "Ответь одним словом. ".
   • We also used instruction data from ShareGPT4V.

The entire training process took 3 days on a single A100 40GB.

Results

The model's performance was evaluated using lmms-eval framework

accelerate launch -m lmms_eval --model llava_hf --model_args pretrained="deepvk/llava-gemma-2b-lora" \
  --tasks gqa-ru,mmbench_ru_dev,gqa,mmbench_en_dev --batch_size 1 \
  --log_samples --log_samples_suffix llava-saiga-8b --output_path ./logs/

Model	GQA	GQA-ru	MMBench	MMBench-ru
deepvk/llava-gemma-2b-lora [this model]	56.39	46.37	51.72	40.19
Intel/llava-gemma-2b	59.80	0.20	39.40	28.30
deepvk/llava-saiga-8b	62.00	51.44	64.26	56.65
llava-hf/llava-1.5-7b-hf	61.31	28.39	62.97	52.25
llava-hf/llava-v1.6-mistral-7b-hf	64.65	6.65	67.70	48.80

Note: for MMBench we didn't use OpenAI API for finding quantifier in generated string. Therefore, the score is similar to Exact Match as in GQA benchmark.

Citation

@misc{liu2023llava,
  title={Visual Instruction Tuning}, 
  author={Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae},
  publisher={NeurIPS},
  year={2023},
}

@misc{deepvk2024llava-gemma-2b-lora,
    title={LLaVA-Gemma-2b-LORA},
    author={Belopolskih, Daniil and Spirin, Egor},
    url={https://huggingface.co./deepvk/llava-gemma-2b-lora},
    publisher={Hugging Face}
    year={2024},
}