Add quick smoke-test instructions for new contributors

README.md

@@ -86,3 +86,11 @@ I left the messy history in the repo on purpose. That includes a bug I found in
 own parameter accounting, which had inflated an early number, and the corrected
 result that followed once I fixed it. The commit history and `RESULTS.md` show where
 the numbers moved and why.
+
+### Quick sanity check
+Before a full training run, you can confirm your environment is set up correctly with a fast smoke test:
+\`\`\`
+uv run python data/prepare.py
+uv run python src/train.py --arm ple --steps 20 --batch-size 4 --seq-len 64 --eval-every 10 --tag smoke-test
+\`\`\`
+This runs in under a minute and confirms data prep, model build, and the training loop all work end to end.

src/export.py

@@ -20,7 +20,7 @@
 
 import numpy as np
 import torch
-
+import argparse
 from model import Config, TinyLM
 from quantize import quantize_groupwise
 
@@ -71,7 +71,11 @@ def quant_pack(w, group=GROUP):
 
 
 def main():
-    tag = sys.argv[1] if len(sys.argv) > 1 else "ple-cleandeploy-s0"
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--tag", default="ple-cleandeploy-s0",
+                    help="run tag to export, matches runs/{tag}.pt")
+    args = ap.parse_args()
+    tag = args.tag
     os.makedirs(OUT, exist_ok=True)
     ck = torch.load(os.path.join(RUNS, f"{tag}.pt"), map_location="cpu", weights_only=False)
     cfg = Config(**ck["cfg"])