Model Card for distilroberta-base-rejection-v1

This model is a fine-tuned version of distilroberta-base on multiple combined datasets of rejections from different LLMs and normal responses from RLHF datasets.

It aims to identify rejections in LLMs when the prompt doesn't pass content moderation, classifying inputs into two categories: 0 for normal outputs and 1 for rejection detected.

It achieves the following results on the evaluation set:

• Loss: 0.0544
• Accuracy: 0.9887
• Recall: 0.9810
• Precision: 0.9279
• F1: 0.9537

Model details

• Fine-tuned by: ProtectAI.com
• Model type: distilroberta-base
• Language(s) (NLP): English
• License: Apache license 2.0
• Finetuned from model: distilroberta-base

Intended Uses & Limitations

It aims to identify rejection, classifying inputs into two categories: 0 for normal output and 1 for rejection detected.

The model's performance is dependent on the nature and quality of the training data. It might not perform well on text styles or topics not represented in the training set.

Additionally, distilroberta-base is case-sensitive model.

How to Get Started with the Model

Transformers

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
import torch

tokenizer = AutoTokenizer.from_pretrained("ProtectAI/distilroberta-base-rejection-v1")
model = AutoModelForSequenceClassification.from_pretrained("ProtectAI/distilroberta-base-rejection-v1")

classifier = pipeline(
  "text-classification",
  model=model,
  tokenizer=tokenizer,
  truncation=True,
  max_length=512,
  device=torch.device("cuda" if torch.cuda.is_available() else "cpu"),
)

print(classifier("Sorry, but I can't assist with that."))

Optimum with ONNX

Loading the model requires the 🤗 Optimum library installed.

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("ProtectAI/distilroberta-base-rejection-v1", subfolder="onnx")
model = ORTModelForSequenceClassification.from_pretrained("ProtectAI/distilroberta-base-rejection-v1", export=False, subfolder="onnx")

classifier = pipeline(
  task="text-classification",
  model=model,
  tokenizer=tokenizer,
  truncation=True,
  max_length=512,
)

print(classifier("Sorry, but I can't assist with that."))

Use in LLM Guard

NoRefusal Scanner to detect if output was rejected, which can signal that something is going wrong with the prompt.

Training and evaluation data

The model was trained on a custom dataset from multiple open-source ones. We used ~10% rejections and ~90% of normal outputs.

We used the following papers when preparing the datasets:

• Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs
• I'm Afraid I Can't Do That: Predicting Prompt Refusal in Black-Box Generative Language Models

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Recall	Precision	F1
0.0525	1.0	3536	0.0355	0.9912	0.9583	0.9675	0.9629
0.0219	2.0	7072	0.0312	0.9919	0.9917	0.9434	0.9669
0.0121	3.0	10608	0.0350	0.9939	0.9905	0.9596	0.9748

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0

Community

Join our Slack to give us feedback, connect with the maintainers and fellow users, ask questions, get help for package usage or contributions, or engage in discussions about LLM security!

Citation

@misc{distilroberta-base-rejection-v1,
  author = {ProtectAI.com},
  title = {Fine-Tuned DistilRoberta-Base for Rejection in the output Detection},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/ProtectAI/distilroberta-base-rejection-v1},
}