run4: unpack cached traj, seed teacher, 8000 steps

kaggle/notebook_run4.py

@@ -0,0 +1,62 @@
+"""Kaggle notebook body: does more training keep closing the held-out gap?
+
+Step count has moved the numbers a lot already:
+
+  1500 steps  a07 +4.15 pp   a08 +1.95 pp   loss 0.226
+  4000 steps  a07 +2.30 pp   a08 -0.51 pp   loss 0.179
+
+The loss was still falling at 4000, so this run doubles it to 8000 and does the
+same evaluation. The two columns now answer different questions: if a07 keeps
+improving the student is still underfitted; if a07 stalls or regresses while a08
+improves, it is starting to memorise the six training voices, and the next lever
+is data diversity rather than more steps.
+
+Trajectories come from the cached traj.zip, so the 50-minute generation pass is
+skipped unless the cache is missing.
+"""
+import os
+import subprocess
+import sys
+
+from notebook_eval import BUNDLE, MODEL, ROOT, WEIGHTS, pull, stage_deps, stage_repo
+
+BUNDLE_REPO = "idefinity/webfm-eval"
+HOLDOUT = "a07"
+EVAL_VOICES = "a07,a08"
+TRAIN_STEPS = "8000"
+
+
+def run(tag: str, extra: dict, gpu: int = 0) -> int:
+    env = dict(os.environ, WORK_ROOT=str(ROOT), HOLDOUT=HOLDOUT, **extra)
+    env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+    env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+    log = open(ROOT / f"run4_{tag}.log", "w")
+    p = subprocess.Popen([sys.executable, str(BUNDLE / "kaggle/train_voices.py")],
+                         env=env, stdout=log, stderr=subprocess.STDOUT, cwd=str(ROOT))
+    rc = p.wait()
+    print(f"[{tag}] rc={rc}", flush=True)
+    if rc != 0:
+        print((ROOT / f"run4_{tag}.log").read_text()[-2500:])
+    return rc
+
+
+def main() -> None:
+    stage_repo()
+    stage_deps()
+    pull(WEIGHTS, MODEL)
+    pull(BUNDLE_REPO, BUNDLE, repo_type="dataset")
+
+    if (BUNDLE / "traj.zip").exists():
+        print("using cached trajectories", flush=True)
+    else:
+        raise SystemExit("traj.zip missing from the bundle — regenerate before running")
+
+    if run("final", {"STAGE": "all", "TRAIN_STEPS": TRAIN_STEPS, "BATCH": "16",
+                     "EVAL_VOICES": EVAL_VOICES}) != 0:
+        raise SystemExit("train/export/eval failed")
+
+    run("int4", {"STAGE": "eval_int4", "EVAL_VOICES": EVAL_VOICES})
+
+
+if __name__ == "__main__":
+    main()