result of causal_tracing_comparison

[matekrk]

Hi!

Thank you for providing the code for your excellent paper.

I wonder how to process the output of causal_tracing_comparison script. It is quite big .json file which is list of lists of dict.
I want either to inspect like you did in figure 5 (actually the same, I want to reproduce the result) or to get the circuit to compute part of networks and see the logit similarites / mismatches and follow it through like in figures 9 and 10.

Many thanks for your help and congratulations again!

[Boshi-Wang]

Thanks for the interest!

For the result, the first dimension is along the model checkpoints, and the second is along the examples. Each element is a dictionary containing the MRRs & causal strengths for different layers & positions, etc.

I added visualization.ipynb which contains the code for processing the results and plotting the figures, if it helps.