Gemma Scope: helping the safety community shed light on the inner workings of language models

Asserting a complete, open suite of sparse autoencoders for language mannequin interpretability.

To create a synthetic intelligence (AI) language mannequin, researchers construct a system that learns from huge quantities of knowledge with out human steering. Because of this, the internal workings of language fashions are sometimes a thriller, even to the researchers who practice them. Mechanistic interpretability is a analysis discipline centered on deciphering these internal workings. Researchers on this discipline use sparse autoencoders as a form of ‘microscope’ that lets them see inside a language mannequin, and get a greater sense of the way it works.

At this time, we’re asserting Gemma Scope, a brand new set of instruments to assist researchers perceive the internal workings of Gemma 2, our light-weight household of open fashions. Gemma Scope is a group of lots of of freely obtainable, open sparse autoencoders (SAEs) for Gemma 2 9B and Gemma 2 2B. We’re additionally open sourcing Mishax, a device we constructed that enabled a lot of the interpretability work behind Gemma Scope.

We hope at present’s launch permits extra formidable interpretability analysis. Additional analysis has the potential to assist the sector construct extra sturdy techniques, develop higher safeguards towards mannequin hallucinations, and shield towards dangers from autonomous AI brokers like deception or manipulation.

Strive our interactive Gemma Scope demo, courtesy of Neuronpedia.

Decoding what occurs inside a language mannequin

If you ask a language mannequin a query, it turns your textual content enter right into a collection of ‘activations’. These activations map the relationships between the phrases you’ve entered, serving to the mannequin make connections between totally different phrases, which it makes use of to write down a solution.

Because the mannequin processes textual content enter, activations at totally different layers within the mannequin’s neural community symbolize a number of more and more superior ideas, often called ‘options’.

For instance, a mannequin’s early layers would possibly study to recall information like that Michael Jordan performs basketball, whereas later layers could acknowledge extra advanced ideas just like the factuality of the textual content.

Nevertheless, interpretability researchers face a key downside: the mannequin’s activations are a mix of many various options. Within the early days of mechanistic interpretability, researchers hoped that options in a neural community’s activations would line up with particular person neurons, i.e., nodes of data. However sadly, in observe, neurons are lively for a lot of unrelated options. Which means that there isn’t a apparent technique to inform which options are a part of the activation.

That is the place sparse autoencoders are available in.

A given activation will solely be a mix of a small variety of options, despite the fact that the language mannequin is probably going able to detecting hundreds of thousands and even billions of them – i.e., the mannequin makes use of options sparsely. For instance, a language mannequin will take into account relativity when responding to an inquiry about Einstein and take into account eggs when writing about omelettes, however most likely gained’t take into account relativity when writing about omelettes.

Sparse autoencoders leverage this reality to find a set of doable options, and break down every activation right into a small variety of them. Researchers hope that the easiest way for the sparse autoencoder to perform this process is to seek out the precise underlying options that the language mannequin makes use of.

Importantly, at no level on this course of can we – the researchers – inform the sparse autoencoder which options to search for. Because of this, we’re in a position to uncover wealthy constructions that we didn’t predict. Nevertheless, as a result of we don’t instantly know the which means of the found options, we search for significant patterns in examples of textual content the place the sparse autoencoder says the function ‘fires’.

Right here’s an instance wherein the tokens the place the function fires are highlighted in gradients of blue in response to their power:

What makes Gemma Scope distinctive

Prior analysis with sparse autoencoders has primarily centered on investigating the internal workings of tiny fashions or a single layer in bigger fashions. However extra formidable interpretability analysis includes decoding layered, advanced algorithms in bigger fashions.

We skilled sparse autoencoders at each layer and sublayer output of Gemma 2 2B and 9B to construct Gemma Scope, producing greater than 400 sparse autoencoders with greater than 30 million realized options in whole (although many options probably overlap). This device will allow researchers to check how options evolve all through the mannequin and work together and compose to make extra advanced options.

Gemma Scope can be skilled with our new, state-of-the-art JumpReLU SAE structure. The unique sparse autoencoder structure struggled to steadiness the dual targets of detecting which options are current, and estimating their power. The JumpReLU structure makes it simpler to strike this steadiness appropriately, considerably decreasing error.

Coaching so many sparse autoencoders was a big engineering problem, requiring plenty of computing energy. We used about 15% of the coaching compute of Gemma 2 9B (excluding compute for producing distillation labels), saved about 20 Pebibytes (PiB) of activations to disk (about as a lot as 1,000,000 copies of English Wikipedia), and produced lots of of billions of sparse autoencoder parameters in whole.

Pushing the sector ahead

In releasing Gemma Scope, we hope to make Gemma 2 the very best mannequin household for open mechanistic interpretability analysis and to speed up the group’s work on this discipline.

To this point, the interpretability group has made nice progress in understanding small fashions with sparse autoencoders and growing related strategies, like causal interventions, computerized circuit evaluation, function interpretation, and evaluating sparse autoencoders. With Gemma Scope, we hope to see the group scale these strategies to fashionable fashions, analyze extra advanced capabilities like chain-of-thought, and discover real-world functions of interpretability equivalent to tackling issues like hallucinations and jailbreaks that solely come up with bigger fashions.