Recursive reasoning dynamics: analysis pipeline, paper drafts, toy models

Failure=more-chaotic (task-general under validity labeling) reduces to convergence/completeness
detection; mechanism (transient chaos vs multistability vs input-induced) under investigation.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

1 files changed, 86 insertions, 0 deletions

diff --git a/eval_directional_ckpts.py b/eval_directional_ckpts.py
new file mode 100644
index 0000000..68653e6
--- /dev/null
+++ b/eval_directional_ckpts.py
@@ -0,0 +1,86 @@
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import sys
+import time
+from pathlib import Path
+
+import torch
+
+
+FLOSS_DIR = Path(__file__).resolve().parent
+sys.path.insert(0, str(FLOSS_DIR))
+
+from step7_interfloss import evaluate, load_model  # noqa: E402
+
+
+def parse_steps(text: str) -> list[int]:
+    return [int(x.strip()) for x in text.split(",") if x.strip()]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--ckpt-root", required=True)
+    parser.add_argument("--steps", required=True)
+    parser.add_argument("--n", type=int, default=1000)
+    parser.add_argument("--batch-size", type=int, default=32)
+    parser.add_argument("--seed", type=int, default=20260611)
+    parser.add_argument("--out", required=True)
+    args = parser.parse_args()
+
+    ckpt_root = Path(args.ckpt_root)
+    rows = []
+    for step in parse_steps(args.steps):
+        ckpt_name = f"step_{step}"
+        if not (ckpt_root / ckpt_name).exists():
+            print(f"[skip] missing {ckpt_name}", flush=True)
+            continue
+
+        t0 = time.time()
+        head, base, cfg, _adam_cls, _sparse_cls = load_model(
+            "trm",
+            ckpt_root,
+            ckpt_name,
+            "cuda",
+            batch_size_override=args.batch_size,
+        )
+        data_path = Path(cfg["data_path"])
+        acc, tok = evaluate(
+            head,
+            base,
+            data_path,
+            n_samples=args.n,
+            batch_size=args.batch_size,
+            device="cuda",
+            seed=args.seed,
+        )
+        elapsed = time.time() - t0
+        row = {
+            "step": step,
+            "n": args.n,
+            "seed": args.seed,
+            "exact_acc": acc,
+            "token_acc": tok,
+            "elapsed_sec": elapsed,
+        }
+        rows.append(row)
+        print(
+            f"step={step} exact={acc:.4f} token={tok:.4f} elapsed={elapsed:.1f}s",
+            flush=True,
+        )
+        del head, base
+        torch.cuda.empty_cache()
+
+        out_path = Path(args.out)
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        out_path.with_suffix(".json").write_text(json.dumps(rows, indent=2))
+        with out_path.open("w", newline="") as f:
+            writer = csv.DictWriter(f, fieldnames=list(rows[0].keys()))
+            writer.writeheader()
+            writer.writerows(rows)
+
+
+if __name__ == "__main__":
+    main()