123 lines
4.6 KiB
Markdown
123 lines
4.6 KiB
Markdown
|
---
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen2.5-Coder-7B-Instruct
|
||
|
|
datasets:
|
||
|
|
- luzimu/WebGen-Bench
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
library_name: transformers
|
||
|
|
license: mit
|
||
|
|
metrics:
|
||
|
|
- accuracy
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- code-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# WebGen-LM
|
||
|
|
|
||
|
|
WebGen-LM is a code language model specifically trained for generating interactive and functional websites from scratch. It is trained using the Bolt.diy trajectories generated from a subset of the training set of WebGen-Bench (🤗 [luzimu/WebGen-Bench](https://huggingface.co/datasets/luzimu/WebGen-Bench)). It has been introduced in the paper [WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch](https://arxiv.org/abs/2505.03733).
|
||
|
|
|
||
|
|
The training data and code can be found at [WebGen-Bench (Github)](https://github.com/mnluzimu/WebGen-Bench).
|
||
|
|
|
||
|
|
The WebGen-LM family of models are as follows:
|
||
|
|
|
||
|
|
|Models | HF Links |
|
||
|
|
|---|---|
|
||
|
|
|WebGen-LM-7B | 🤗 [luzimu/WebGen-LM-7B](https://huggingface.co/luzimu/WebGen-LM-7B) |
|
||
|
|
|WebGen-LM-14B | 🤗 [luzimu/WebGen-LM-14B](https://huggingface.co/luzimu/WebGen-LM-14B) |
|
||
|
|
|WebGen-LM-32B | 🤗 [luzimu/WebGen-LM-32B](https://huggingface.co/luzimu/WebGen-LM-32B) |
|
||
|
|
|
||
|
|
## Performance on WebGen-Bench
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
You can use `WebGen-LM` with the `transformers` library to generate website code.
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
import torch
|
||
|
|
|
||
|
|
model_id = "luzimu/WebGen-LM-32B" # You can also use WebGen-LM-7B or WebGen-LM-14B
|
||
|
|
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
model_id,
|
||
|
|
torch_dtype=torch.bfloat16,
|
||
|
|
device_map="auto"
|
||
|
|
)
|
||
|
|
|
||
|
|
# Example for website generation
|
||
|
|
prompt = """Generate the complete HTML, CSS, and JavaScript code for a responsive website.
|
||
|
|
The website should be a simple landing page for a coffee shop.
|
||
|
|
It needs:
|
||
|
|
1. A navigation bar at the top with "Home", "Menu", "About Us", and "Contact" links.
|
||
|
|
2. A hero section with a background image, a title "Brewing Perfection", and a call-to-action button "View Our Menu".
|
||
|
|
3. A menu section displaying at least 3 coffee items with their names and prices.
|
||
|
|
4. An "About Us" section with a brief description of the coffee shop.
|
||
|
|
5. A "Contact" section with an address, phone number, and a simple contact form (Name, Email, Message, Submit button).
|
||
|
|
6. Basic responsive design for mobile views.
|
||
|
|
"""
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "user", "content": prompt}
|
||
|
|
]
|
||
|
|
|
||
|
|
text = tokenizer.apply_chat_template(
|
||
|
|
messages,
|
||
|
|
tokenize=False,
|
||
|
|
add_generation_prompt=True
|
||
|
|
)
|
||
|
|
|
||
|
|
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
|
|
|
||
|
|
generated_ids = model.generate(
|
||
|
|
model_inputs.input_ids,
|
||
|
|
max_new_tokens=2048, # Adjust as needed for full website code
|
||
|
|
do_sample=True,
|
||
|
|
temperature=0.7,
|
||
|
|
top_p=0.9,
|
||
|
|
repetition_penalty=1.05,
|
||
|
|
)
|
||
|
|
|
||
|
|
# Decode the generated output, skipping special tokens
|
||
|
|
response = tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)[0]
|
||
|
|
# The response will contain the full conversation history including the input prompt.
|
||
|
|
# To get only the newly generated text, you might need to slice it or use the appropriate
|
||
|
|
# tokenizer behavior based on how apply_chat_template adds prompt.
|
||
|
|
# For simplicity, if the model just appends to the prompt, direct decode might suffice.
|
||
|
|
# A more robust approach might be:
|
||
|
|
# generated_text_only = tokenizer.decode(generated_ids[0][len(model_inputs.input_ids[0]):], skip_special_tokens=True)
|
||
|
|
print(response)
|
||
|
|
|
||
|
|
# You might need to parse the output to separate HTML, CSS, and JS if the model outputs a combined file.
|
||
|
|
# For example, look for specific markers like <html>, <style>, <script>
|
||
|
|
```
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
If you find our project useful, please cite:
|
||
|
|
|
||
|
|
```
|
||
|
|
@misc{lu2025webgenbenchevaluatingllmsgenerating,
|
||
|
|
title={WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch},
|
||
|
|
author={Zimu Lu and Yunqiao Yang and Houxing Ren and Haotian Hou and Han Xiao and Ke Wang and Weikang Shi and Aojun Zhou and Mingjie Zhan and Hongsheng Li},
|
||
|
|
year={2025},
|
||
|
|
eprint={2505.03733},
|
||
|
|
archivePrefix={arXiv},
|
||
|
|
primaryClass={cs.CL},
|
||
|
|
url={https://arxiv.org/abs/2505.03733},
|
||
|
|
}
|
||
|
|
|
||
|
|
@misc{lu2025webgenagentenhancinginteractivewebsite,
|
||
|
|
title={WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning},
|
||
|
|
author={Zimu Lu and Houxing Ren and Yunqiao Yang and Ke Wang and Zhuofan Zong and Junting Pan and Mingjie Zhan and Hongsheng Li},
|
||
|
|
year={2025},
|
||
|
|
eprint={2509.22644},
|
||
|
|
archivePrefix={arXiv},
|
||
|
|
primaryClass={cs.CL},
|
||
|
|
url={https://arxiv.org/abs/2509.22644},
|
||
|
|
}
|
||
|
|
```
|