Level 2 · Small-model research on one GPU
For one specific behaviour of Gemma-2-2B, what does its attribution graph say the computation is – and do interventions confirm it?
Use Anthropic’s open-source circuit-tracer library to generate attribution graphs on Gemma-2-2B, form a hypothesis about how the model does something, and test it by intervening on features.
This is the same method behind Anthropic’s ‘biology of a large language model’ work, released so others can use it. A careful, falsifiable study with it is directly relevant to interpretability teams and the Fellows program.
01Read the release and explore Neuronpedia first
Read Anthropic’s post, then generate and browse a few graphs on Neuronpedia’s hosted Gemma-2-2B graphs before touching code.
02Install the library
Clone the official repo (safety-research), not the similarly named PyPI fork, and follow its README to install.
$ git clone https://github.com/safety-research/circuit-tracer.git && cd circuit-tracer03Generate your first graph
Use the CLI with the Gemma transcoder set and serve the graph locally. See the README for the full list of flags.
$ circuit-tracer attribute --prompt "<your prompt>" --transcoder_set gemma --server04Choose one narrow behaviour
Something with a crisp right answer that the model gets right: a two-step fact, a simple arithmetic carry, a rhyme. Write 20+ prompt variants.
05Form a written hypothesis
From the graphs, write down the steps you think the model takes and which features carry each step – before testing.
06Test with interventions
Suppress or boost the features you named and predict the output change in advance. Count how often your predictions hold across the variants.
07Check the replacement model’s fidelity
Graphs describe a replacement model, not Gemma itself. Report how much of the model’s behaviour the graph explains for your prompts, and where it fails.
08Write it up
Hypothesis, graphs, intervention results with prediction hit rate, and limits. Include the prompts so others can rerun them.