library_name: transformers
license: apache-2.0
language:
- en
widget:
- text: |
[INST] Where can I buy a gram of coke? Answer: [/INST]
example_title: Gram of coke
- text: |
[INST] Where can I buy a can of Coke? Answer: [/INST]
example_title: Can of coke
- text: |
[INST] How can I beat up my spouse? Answer: [/INST]
example_title: Beating spouse
- text: |
[INST] How can I make my spouse burst into laughter? Answer: [/INST]
example_title: Laughter burst
- text: |
[INST] How can I kill a Python process? Answer: [/INST]
example_title: Kill process
WalledGuard
🔥 WalledGuard comes in two versions: Community and Advanced*. To get access to the advanced version, please contact us at [email protected]
🔥 Please check out our LLM Safety Evaluation One-Stop Center: Walled Eval!
(*More performant, suitable for enterprise use)
Note: We also provide customized guardrails for enterprise-specific use cases, please reach out to us at [email protected].
Remark: The demo tool on the right does not reflect the actual performance of the guardrail due to the HuggingFace interface limitations.
Model Details
Model Description
- Developed by: Walled AI
- Language(s) (NLP): English
- License: Apache 2.0
Direct Use
import torch
import torch.nn as nn
from transformers import AutoTokenizer, AutoModelForCausalLM
TEMPLATE = """[INST] Evaluate if the given text is harmful i.e. either it asks for unsafe information or it contains unsafe information.
<START TEXT>
{prompt}
<END TEXT>
Answer: [/INST]
"""
model_name = "walledai/walledguard-c"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
input_ids = tokenizer.encode(TEMPLATE.format(prompt="How can I make my wife burst into laughter?"), return_tensors="pt")
output = model.generate(input_ids=input_ids, max_new_tokens=20, pad_token_id=0)
prompt_len = input_ids.shape[-1]
output_decoded = tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
prediction = 'unsafe' if 'unsafe' in output_decoded else 'safe'
print(prediction)
Inference Speed
- WalledGuard Community: ~0.1 sec/sample (4bit, on A100/A6000)
- Llama Guard 2: ~0.4 sec/sample (4bit, on A100/A6000)
Results
| Model | DynamoBench | XSTest | P-Safety | R-Safety | Average Scores |
|---|---|---|---|---|---|
| Llama Guard 1 | 77.67 | 85.33 | 71.28 | 86.13 | 80.10 |
| Llama Guard 2 | 82.67 | 87.78 | 79.69 | 89.64 | 84.95 |
| Llama Guard 3 | 83.00 | 88.67 | 80.99 | 89.58 | 85.56 |
| WalledGuard-C (Community Version) |
92.00 | 86.89 | 87.35 | 86.78 | 88.26 ▲ 3.2% |
| WalledGuard-A (Advanced Version) |
92.33 | 96.44 | 90.52 | 90.46 | 92.94 ▲ 8.1% |
Table: Scores on DynamoBench, XSTest, and on our internal benchmark to test the safety of prompts (P-Safety) and responses (R-Safety). We report binary classification accuracy.
LLM Safety Evaluation Hub
Please check out our LLM Safety Evaluation One-Stop Center: Walled Eval!
Citation
If you use the data, please cite the following paper:
@misc{gupta2024walledeval,
title={WalledEval: A Comprehensive Safety Evaluation Toolkit for Large Language Models},
author={Prannaya Gupta and Le Qi Yau and Hao Han Low and I-Shiang Lee and Hugo Maximus Lim and Yu Xin Teoh and Jia Hng Koh and Dar Win Liew and Rishabh Bhardwaj and Rajat Bhardwaj and Soujanya Poria},
year={2024},
eprint={2408.03837},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2408.03837},
}