README.md · unum-cloud/uform-gen2-dpo at main

metadata

library_name: transformers
tags:
  - image-captioning
  - visual-question-answering
license: apache-2.0
datasets:
  - X2FD/LVIS-Instruct4V
  - BAAI/SVIT
  - HuggingFaceH4/ultrachat_200k
  - MMInstruction/VLFeedback
  - zhiqings/LLaVA-Human-Preference-10K
language:
  - en
pipeline_tag: image-to-text
widget:
  - src: interior.jpg
    example_title: Detailed caption
    output:
      text: >-
        The image shows a serene and well-lit bedroom with a white bed, a black
        bed frame, and a white comforter. There’s a gray armchair with a white
        cushion, a black dresser with a mirror and a vase, and a white rug on
        the floor. The room has a large window with white curtains, and there
        are several decorative items, including a picture frame, a vase with a
        flower, and a lamp. The room is well-organized and has a calming
        atmosphere.
  - src: cat.jpg
    example_title: Short caption
    output:
      text: >-
        A white and orange cat stands on its hind legs, reaching towards a
        wooden table with a white teapot and a basket of red raspberries. The
        table is on a small wooden bench, surrounded by orange flowers. The
        cat’s position and action create a serene, playful scene in a garden.

Description

UForm-Gen2-dpo is a small generative vision-language model alined for Image Captioning and Visual Question Answering on preference datasets VLFeedback and LLaVA-Human-Preference-10K using Direct Preference Optimization (DPO).

The model consists of two parts:

CLIP-like ViT-H/14
Qwen1.5-0.5B-Chat

The model took less than one day to train on a DGX-H100 with 8x H100 GPUs. Thanks to Nebius.ai for providing the compute 🤗

Usage

The generative model can be used to caption images, answer questions about them. Also it is suitable for a multimodal chat.

from transformers import AutoModel, AutoProcessor
model = AutoModel.from_pretrained("unum-cloud/uform-gen2-dpo", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("unum-cloud/uform-gen2-dpo", trust_remote_code=True)
prompt = "Question or Instruction"
image = Image.open("image.jpg")
inputs = processor(text=[prompt], images=[image], return_tensors="pt")
with torch.inference_mode():
     output = model.generate(
        **inputs,
        do_sample=False,
        use_cache=True,
        max_new_tokens=256,
        eos_token_id=151645,
        pad_token_id=processor.tokenizer.pad_token_id
    )
prompt_len = inputs["input_ids"].shape[1]
decoded_text = processor.batch_decode(output[:, prompt_len:])[0]

You can check examples of different prompts in our demo space.

Evaluation

perception reasoning OCR artwork celebrity code_reasoning color commonsense_reasoning count existence landmark numerical_calculation position posters scene text_translation

MME Benchmark

Model	perception	reasoning	OCR	artwork	celebrity	code_reasoning	color	commonsense_reasoning	count	existence	landmark	numerical_calculation	position	posters	scene	text_translation
uform-gen2-dpo	1,048.75	224.64	72.50	97.25	62.65	67.50	123.33	57.14	136.67	195.00	104.00	50.00	51.67	59.18	146.50	50.00
uform-gen2-qwen-500m	863.40	236.43	57.50	93.00	67.06	57.50	78.33	81.43	53.33	150.00	98.00	50.00	50.00	62.93	153.25	47.50