metadata

base_model: unsloth/Llama-3.2-90B-Vision-Instruct-bnb-4bit
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - mllama
license: apache-2.0
language:
  - en

Reverse Browser 90B Model

License: Apache-2.0
Finetuned from model : unsloth/Llama-3.2-90B-Vision-Instruct-bnb-4bit

This is a vector-to-ui model. It was fine-tuned on the HTML+CSS code of hundreds of thousands of public web pages and their SVG counterparts.

The goal of the model is to aid rapid prototyping and web development.

The model was trained on mobile resolution (393×852) SVGs only. Its test set accuracy was 0.3012 LPIPS (AlexNet). This suggests that it is not ready for industrial use in a real workflow, but it may serve research purposes.

Usage

!pip install vllm

from vllm import LLM, SamplingParams
llm = LLM(
    "tcz/rb-llama-90b",  
    tensor_parallel_size=4,   # The was tested on four H100s (96GB) GPUs
    dtype="bfloat16",
    gpu_memory_utilization=0.9, 
    enforce_eager=True,         
    max_model_len=12_000,
    max_num_seqs=64,
)

data_prompt = """Your job is to take an SVG file of a web design and convert it into a pixel-perfect HTML and CSS markup and stylesheet.

### Input:
{}

### Response:
{}"""

max_tokens = 12_000

# Experiment with different temperature and top-p settings
sampling_params = SamplingParams(
    temperature   = 0.0,
    top_p         = 1.0,
    top_k         = 1,
    n             = 1,
    max_tokens    = max_tokens,
    seed          = 0
)

prompt = data_prompt.format(
    YOUR_SVG_CONTENT,
    "",
)

output = llm.generate([prompt], sampling_params)
print(output[0].outputs[0].text)