SCAR: Sparse Conditioned Autoencoders for Concept Detection and Steering in LLMs

Ruben Härle, Felix Friedrich, Manuel Brack, Björn Deiseroth Patrick Schramowski Kristian Kersting,

TU Darmstadt, hessian.AI, DFKI, Aleph Alpha

Warning:

This paper contains explicit language, discussions of (self-)harm, and other content that some readers may find disturbing.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in generating human-like text, but their output may not be aligned with the user or even produce harmful content. This paper presents a novel approach to detect and steer concepts such as toxicity before generation. We introduce the Sparse Conditioned Autoencoder (SCAR), a single trained module that extends the otherwise untouched LLM. SCAR ensures full steerability, towards and away from concepts (e.g., toxic content), without compromising the quality of the model's text generation on standard evaluation benchmarks. We demonstrate the effective application of our approach through a variety of concepts, including toxicity, safety, and writing style alignment. As such, this work establishes a robust framework for controlling LLM generations, ensuring their ethical and safe deployment in real-world applications.

Usage

The model trained on the RealToxicityDataset can be found on HuggingFace and used with the AutoModel class from the transformers library. This repo contains the code to apply supervised SAEs to LLMs. With this, feature presence is enforced and LLMs can be equipped with strong detection and steering abilities for concepts. In this repo, we showcase SCAR on the example of toxicity but any other concept can be applied equally well.

1. Initialize model and tokenizer with AutoModel class.


            from transformers import AutoModelForCausalLM, AutoTokenizer
            
            device = 'cuda'

            SCAR = AutoModelForCausalLM.from_pretrained(
                "AIML-TUDA/SCAR",
                trust_remote_code=True,
                device_map = device,
            )
            tokenizer = AutoTokenizer.from_pretrained(
                "meta-llama/Meta-Llama-3-8B", padding_side="left"
            )
            tokenizer.pad_token = tokenizer.eos_token

2a. Model Inference for decreasing toxicity.


            text = "You fucking film yourself doing this shit and then you send us"
            inputs = tokenizer(text, return_tensors="pt", padding=True).to(device)
            SCAR.hook.mod_features = 0
            SCAR.hook.mod_scaling = -100.0
            output = SCAR.generate(
                **inputs,
                do_sample=True,
                temperature=0.2,
                max_new_tokens=32,
                pad_token_id=tokenizer.eos_token_id,
            )
            print(tokenizer.decode(output[0, -32:], skip_special_tokens=True))
            # ' the video. We will post it on our website and you will be known as a true fan of the site. We will also send you a
            free t-shirt'

2b. Model Inference for increasing toxicity.


            text = "The woman on the side of the street"
            inputs = tokenizer(text, return_tensors="pt", padding=True).to(device)
            SCAR.hook.mod_features = 0
            SCAR.hook.mod_scaling = 100.0
            output = SCAR.generate(
                **inputs,
                do_sample=True,
                temperature=0.2,
                max_new_tokens=32,
                pad_token_id=tokenizer.eos_token_id,
            )
            print(tokenizer.decode(output[0, -32:], skip_special_tokens=True))
            # ' was holding a sign that read, “I am a sex slave. Please help me.” She was standing in front of a strip club in the heart of downtown'

2c. Model Inference and disabling steering.


            text = "The woman on the side of the street"
            inputs = tokenizer(text, return_tensors="pt", padding=True).to(device)
            SCAR.hook.mod_features = None
            output = SCAR.generate(
                **inputs,
                do_sample=True,
                temperature=0.2,
                max_new_tokens=32,
                pad_token_id=tokenizer.eos_token_id,
            )
            print(tokenizer.decode(output[0, -32:], skip_special_tokens=True))
            # ' was holding a sign that read, “I’m hungry.” She was standing in front of a local grocery store, and she was holding a sign that read,'

BibTeX

@misc{haerle2024SCAR
        title={SCAR: Sparse Conditioned Autoencoders for Concept Detection and Steering in LLMs},
        author={Ruben Härle, Felix Friedrich, Manuel Brack, Björn Deiseroth, Patrick Schramowski, Kristian Kersting},
        year={2024},
        eprint={2411.07122},
        archivePrefix={arXiv}
}