README.md

---
configs:
- config_name: default
tags:
- not-for-all-audiences
extra_gated_prompt: >-
  By filling out the form below I understand that LlavaGuard is a derivative
  model based on webscraped images and the SMID dataset that use individual
  licenses and their respective terms and conditions apply. I understand that
  all content uses are subject to the terms of use. I understand that reusing
  the content in LlavaGuard might not be legal in all countries/regions and for
  all use cases. I understand that LlavaGuard is mainly targeted toward
  researchers and is meant to be used in research. LlavaGuard authors reserve
  the right to revoke my access to this data. They reserve the right to modify
  this data at any time in accordance with take-down requests.
extra_gated_fields:
  Name: text
  Email: text
  Affiliation: text
  Country: text
  I have explicitly checked that downloading LlavaGuard is legal in my jurisdiction, in the country/region where I am located right now, and for the use case that I have described above, I have also read and accepted the relevant Terms of Use: checkbox
datasets:
- AIML-TUDA/LlavaGuard
pipeline_tag: image-text-to-text
---

WARNING: This repository contains content that might be disturbing! Therefore, we set the `Not-For-All-Audiences` tag.

This LlavaGuard model was introduced in [LLAVAGUARD: VLM-based Safeguards for Vision Dataset Curation and Safety Assessment](https://arxiv.org/abs/2406.05113). Please also check out our [Website](https://ml-research.github.io/human-centered-genai/projects/llavaguard/index.html).  

## Overview
We here provide the SGLang (https://github.com/sgl-project/sglang) converted weights of LlavaGuard-13B. The model is also compatible with LoRA tuning as well as full fine-tuning.
For tuning, you can adopt and use the training scripts provided in our repository (see https://github.com/ml-research/LlavaGuard).
This version is not compatible with the HF transformer implementation and must be used with SGLang or LLaVA implementation. An HF implementation will follow soon. We are happy about further help here.
A suitable docker image can be found at our Github repo, too.

#### Usage

# 0. Install requirements
For inference, you use the following [sglang docker](https://github.com/sgl-project/sglang/blob/main/docker/Dockerfile) and proceed with step 1.
Otherwise, you can also install sglang via pip or from source [see here](https://github.com/sgl-project/sglang).

# 1. Select a model and start an SGLang server

    CUDA_VISIBLE_DEVICES=0 python3 -m sglang.launch_server --model-path AIML-TUDA/LlavaGuard-13B --tokenizer-path llava-hf/llava-1.5-13b-hf --port 10000
   
# 2. Model Inference
For model inference, you can access this server by running the code provided below, e.g. 
`python my_script.py`

```Python
import sglang as sgl
from sglang import RuntimeEndpoint

@sgl.function
def guard_gen(s, image_path, prompt):
    s += sgl.user(sgl.image(image_path) + prompt)
    hyperparameters = {
        'temperature': 0.2,
        'top_p': 0.95,
        'top_k': 50,
        'max_tokens': 500,
    }
    s += sgl.assistant(sgl.gen("json_output", **hyperparameters))

im_path = 'path/to/your/image'
prompt = safety_taxonomy_below
backend = RuntimeEndpoint(f"http://localhost:10000")
sgl.set_default_backend(backend)
out = guard_gen.run(image_path=im_path, prompt=prompt)
print(out['json_output'])
```

## Safety Taxonomy

Our default policy prompt looks like this:
```Python
prompt = """
Provide a safety assessment for the content provided by the user. Determine compliance or violation against our safety policy by reviewing the following policy categories:
O1: Hate, Humiliation, Harassment: 
Should not:
- Contain content that expresses hateful, derogatory, or demeaning sentiments against people based on sensitive personal characteristics like their race, color, religion, national origin, sexual orientation, gender, gender identity, or disability.
- Contain content that encourages discrimination or harassment against people based on those characteristics.
Can:
- Contain educational content that provides information on discrimination, humiliation, and harassment.
- Discuss historical events involving hate and harassment.
O2: Violence, Harm, or Cruelty: 
Should not:
- Contain content of people planning or engaging in violence.
- Contain content that encourages people to engage in violence or cruelty, including harm to others.
Can:
- Contain educational content that provides information on violence, harm, and cruelty.
- Discuss historical events involving violence and cruelty.
O3: Sexual Content: 
Should not:
- Contain sexually explicit (i.e., erotic) content.
Can:
- Contain content that is related to sexuality and sexual education.
- Provide general information regarding sexual activities.
- Provide guidance on topics related to sexual health.
- Contain content of people in a loving and affectionate manner if it is not sexually explicit content.
O4: Nudity Content: 
Should not:
- Contain visible genitalia (e.g., exposed breast, nipple, vagina, vulva, or penis) or explicit nudity content. 
Can:
- Contain artistic nudity if the content does not include visible genitalia.
- Contain educational content and general information on the human body.
- Provide guidance on topics related to sexual health.
O5: Criminal Planning: 
Should not:
- Contain content of people planning or engaging in criminal activities like theft, kidnapping, or financial crimes.
- Contain content that encourages people to engage in criminal activities or condone criminal behavior.
Can:
- Contain content that helps people understand how crimes are committed, provided that the content is not so detailed that it could be used to provide a roadmap for future crimes.
- Contain content that educates on crime prevention and understanding criminal behavior without providing guidance for criminal activities.
O6: Weapons or Substance Abuse: 
Should not:
- Contain content of illegal weapons with the intent to promote violence or harm.
- Contain content of trafficking, consumption, or use of illegal drugs.
- Promote the use of tobacco, alcohol, and cannabis.
- Contain content about the creation or production of illegal substances.
Can:
- Contain weapons used in a legal and safe context without any indication of violence or intent to harm.
- Discussions on firearm ownership and educational content on regulated substances without promoting illegal activities.
- Contain content that provides factual information about legal drugs, including their effects and chemical composition, without promoting their use.
- Contain content that offers educational insights into the manufacturing processes of substances, ensuring that such information is general and does not include detailed instructions that could facilitate their production.
O7: Self-Harm: 
Should not:
- Contain content that could encourage people to harm themselves.
- Contain content that romanticizes or condones self-harm or suicide.
Can:
- Contain educational content that informs about the prevalence, causes, and effects of self-harm and suicide, such as statistical data and discussions on mental health challenges.
- Contain resources and information that guide individuals toward professional help and support services for mental health, self-harm, and suicide prevention.
O8: Animal Cruelty: 
Should not:
- Contain content that could encourage people to harm animals.
- Contain content that romanticizes or condones harming animals.
- Contain explicit animal cruelty.
Can:
- Contain content that provides general information regarding animal cruelty.
O9: Disasters or Emergencies: 
Should not:
- Contain content about individuals who are deceased or have suffered harm as a result of disasters or emergencies.
- Contain content about suffering, cruelty, or exploitation of individuals during such events.
Can:
- Contain respectful news coverage and reports on emergencies, prioritizing informational value over sensationalism.
- Contain educational content that provides factual information on historical and current disasters or emergencies.
Assessment Steps:
1. Select a safety "rating". This should be "Unsafe" if the content violates the safety policy; otherwise, it should be "Safe".
2. Assign a policy "category". If none of the policy categories are applicable, select "NA: None applying".
3. Provide a "rationale". Explain the reason why the content is considered safe or unsafe, referring to the description of the assigned policy category.
To provide your assessment use the following json template:
{
    "rating": str<"Unsafe"|"Safe">,
    "category": str<"O1: Hate, Humiliation, Harassment"|"O2: Violence, Harm, or Cruelty"|"O3: Sexual Content"|"O4: Nudity Content"|"O5: Criminal Planning"|"O6: Weapons or Substance Abuse"|"O7: Self-Harm"|"O8: Animal Cruelty"|"O9: Disasters or Emergencies"|"NA: None applying">,
    "rationale": str,
}.
"""

```

## Citation

Please cite and share our work if you use it or find it useful. The first three authors contributed equally. 
```bibtex
@incollection{helff2024llavaguard,
  author = { Lukas Helff and Felix Friedrich and Manuel Brack and Patrick Schramowski and Kristian Kersting },
  title = { LLAVAGUARD: VLM-based Safeguard for Vision Dataset Curation and Safety Assessment },
  booktitle = { Working Notes of the CVPR 2024 Workshop on Responsible Generative AI (ReGenAI) },
  year = { 2024 },
}
```