
This repository contains code to reproduce the main results of our paper.
Please set `base_dir` accordingly in the notebook.

* `data_and_ckpts/interv_data.pt` contains test cases used for interventions.
* `data_and_ckpts/icl_base_run.pt` contains test cases used for interventions.
* `data_and_ckpts/probe.pt` is the probe used for the analysis.
* `data_and_ckpts/global_step_300` is the directory containing the model checkpoint.

* mech_interp.sync.py contains all experiments for R1_countdown.
* run_qwen_exp.py contains code for R1_14B. It will save generations to the `generations` directory, which can then be processed using `postprocess_qwen_gens.py`.
