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

Saying 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. In consequence, the interior workings of language fashions are sometimes a thriller, even to the researchers who prepare them. Mechanistic interpretability is a analysis discipline targeted on deciphering these interior 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 the moment, we’re announcing Gemma Scope, a brand new set of instruments to assist researchers perceive the interior workings of Gemma 2, our light-weight household of open fashions. Gemma Scope is a set of tons 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 as we speak’s launch allows extra bold interpretability analysis. Additional analysis has the potential to assist the sphere construct extra strong techniques, develop higher safeguards towards mannequin hallucinations, and shield towards dangers from autonomous AI brokers like deception or manipulation.

Try 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 sequence 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 put in writing a solution.

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

For instance, a mannequin’s early layers would possibly study to recall facts like that Michael Jordan plays basketball, whereas later layers might acknowledge extra complicated ideas like the factuality of the text.

Nonetheless, 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 energetic for a lot of unrelated options. Which means there isn’t a apparent option to inform which options are a part of the activation.

That is the place sparse autoencoders are available.

A given activation will solely be a mix of a small variety of options, although 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 contemplate relativity when responding to an inquiry about Einstein and contemplate eggs when writing about omelettes, however most likely gained’t contemplate relativity when writing about omelettes.

Sparse autoencoders leverage this truth to find a set of doable options, and break down every activation right into a small variety of them. Researchers hope that one of the best ways for the sparse autoencoder to perform this process is to search 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. In consequence, we’re in a position to uncover wealthy buildings that we didn’t predict. Nonetheless, as a result of we don’t instantly know the which means of the found options, we search for meaningful 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 based on their power:

What makes Gemma Scope distinctive

Prior analysis with sparse autoencoders has primarily targeted on investigating the interior workings of tiny models or a single layer in larger models. However extra bold interpretability analysis includes decoding layered, complicated 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 complete (although many options seemingly overlap). This device will allow researchers to review how options evolve all through the mannequin and work together and compose to make extra complicated options.

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

Coaching so many sparse autoencoders was a big engineering problem, requiring loads 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 a million copies of English Wikipedia), and produced tons of of billions of sparse autoencoder parameters in complete.

Pushing the sphere ahead

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

To date, the interpretability group has made nice progress in understanding small fashions with sparse autoencoders and growing related methods, like causal interventions, automatic circuit analysis, feature interpretation, and evaluating sparse autoencoders. With Gemma Scope, we hope to see the group scale these methods to trendy fashions, analyze extra complicated capabilities like chain-of-thought, and discover real-world purposes of interpretability similar to tackling issues like hallucinations and jailbreaks that solely come up with bigger fashions.