Upload folder using huggingface_hub

train/curate_pivot_set.py

@@ -122,6 +122,25 @@ def cmd_baseline(args: argparse.Namespace) -> None:
     correct = sum(1 for v in out.values() if v["predicted"] == v["gold"])
     print(f"[baseline] saved to {args.out}. Text-only accuracy: {correct}/{len(out)} = {correct/len(out):.2%}")
 
+    # Auto-upload to HF Space so the next step (filter) can pull it.
+    # Disabled if NO_HF_UPLOAD=1 or no token available.
+    import os
+    if not os.environ.get("NO_HF_UPLOAD") and os.environ.get("HF_TOKEN"):
+        try:
+            from huggingface_hub import HfApi
+            repo_id = os.environ.get("HF_REPO_ID", "aamrinder/subtext-arena")
+            HfApi().upload_file(
+                path_or_fileobj=str(args.out),
+                path_in_repo="data/baseline.json",
+                repo_id=repo_id,
+                repo_type="space",
+                token=os.environ["HF_TOKEN"],
+                commit_message="add text-only Qwen baseline (curate_pivot_set baseline)",
+            )
+            print(f"[baseline] uploaded to {repo_id}/data/baseline.json")
+        except Exception as e:
+            print(f"[baseline] upload failed (saved locally only): {e}")
+
 
 # ---------------------------------------------------------------------------
 # Step 2: filter — confidently-wrong clips are the Pivot candidates