Add anonymous KAFT MVP reproduction

1 files changed, 53 insertions, 0 deletions

diff --git a/run_mvp.py b/run_mvp.py
new file mode 100644
index 0000000..c55b33c
--- /dev/null
+++ b/run_mvp.py
@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+"""Run the anonymous KAFT MVP from the command line."""
+
+from __future__ import annotations
+
+import argparse
+
+import pandas as pd
+
+from kaft_mvp import MVPConfig, run_mvp
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cpu")
+    parser.add_argument("--seeds", default="0,1,2")
+    parser.add_argument("--epochs", type=int, default=200)
+    parser.add_argument("--diagnostic-epochs", type=int, default=100)
+    parser.add_argument("--output-dir", default="artifacts")
+    args = parser.parse_args()
+    seeds = tuple(int(value) for value in args.seeds.split(","))
+    config = MVPConfig(
+        seeds=seeds,
+        epochs=args.epochs,
+        diagnostic_epochs=args.diagnostic_epochs,
+        device=args.device,
+    )
+    payload = run_mvp(config=config, output_dir=args.output_dir)
+    print("\nBP versus KAFT")
+    print(pd.DataFrame(payload["summary"]).to_string(index=False))
+    diagnostic = pd.DataFrame(
+        [
+            {
+                "seed": row["seed"],
+                "all_weight_grads_zero": row[
+                    "all_weight_gradients_exact_zero"
+                ],
+                "output_adjacent_error": row[
+                    "output_adjacent_error_frobenius"
+                ],
+                "hidden_probe_percent": 100.0
+                * row["standardized_penultimate_probe_accuracy"],
+            }
+            for row in payload["gradient_diagnostics"]
+        ]
+    )
+    print("\n10-layer BP diagnostic")
+    print(diagnostic.to_string(index=False))
+    print(f"\nArtifacts written to {args.output_dir}/")
+
+
+if __name__ == "__main__":
+    main()