Dataset Viewer
Auto-converted to Parquet Duplicate
train_shape
dict
val_shape
dict
test_shape
dict
gate_names
listlengths
1
1
{ "x": [ 65536, 2 ], "y": [ 65536, 1 ] }
{ "x": [ 8192, 2 ], "y": [ 8192, 1 ] }
{ "x": [ 8192, 2 ], "y": [ 8192, 1 ] }
[ "XOR" ]
{ "x": [ 65536, 2 ], "y": [ 65536, 1 ] }
{ "x": [ 8192, 2 ], "y": [ 8192, 1 ] }
{ "x": [ 8192, 2 ], "y": [ 8192, 1 ] }
[ "XOR" ]
{ "x": [ 1024, 2 ], "y": [ 1024, 1 ] }
{ "x": [ 128, 2 ], "y": [ 128, 1 ] }
{ "x": [ 128, 2 ], "y": [ 128, 1 ] }
[ "XOR" ]

Which Circuit is it? — experiment data

Run artifacts for the Which Circuit is it? sequence and its parameter-decomposition follow-up. Code: unrulyabstractions/identifiability-toy-study.

Tiny MLPs (2 → 3 → 3 → 1, tanh) trained on XOR. All 48 node patterns are enumerated and scored for observational, interventional, and counterfactual faithfulness. VPD (adVersarial Parameter Decomposition) is then run on the trained models to test what parameter-space methods recover.

Contents

runs/lw_scenarioA/    # trained WITH input noise (sigma = 0.2)
runs/lw_scenarioB/    # trained without noise (sigma = 0.001)
runs/run_20260811-221049-393085/   # fast-test run incl. VPD pipeline artifacts

Each run contains per-trial model.pt, config/metrics JSON, per-node-pattern faithfulness summaries and figures, and (Scenario A) vpd_recovery/ with the VPD vs no-adversary ablation results.

Headline numbers

  • Scenario A: perfect canonical node masks are exactly {34, 44}; interventions and counterfactuals (completeness/independence 0.75 vs ≤ 0.625) pick #34.
  • Scenario B: perfect masks are {43, 44, 47}; the full model deviates from the ideal XOR boundary away from the corners.
  • VPD recovers an input-conditional case analysis (per-input masks land on enumerated patterns, e.g. #47/#46/#34) whose union is the full circuit.
  • Without the adversarial term, PGD reconstruction error grows ~200x and CI-masked reconstruction breaks.

Regenerate any analysis from the saved models with uv run main.py --load-mlp runs/<name> in the code repo.

Downloads last month
45