fix(dataset): enable trust_remote_code for BigBio; robust load_dataset retry

app.py

@@ -64,12 +64,15 @@ def _train_ner_lora(
         if lower_spec in alias_map:
             ds_name, ds_config = alias_map[lower_spec]
             log(f"Using alias mapping: {ds_spec} -> {ds_name}:{ds_config}")
-            ds = load_dataset(ds_name, ds_config)
+            ds = load_dataset(ds_name, ds_config, trust_remote_code=True)
         elif ":" in ds_spec:
             ds_name, ds_config = [s.strip() for s in ds_spec.split(":", 1)]
-            ds = load_dataset(ds_name, ds_config)
+            # If loading from community repo, allow remote code
+            trust = "/" in ds_name
+            ds = load_dataset(ds_name, ds_config, trust_remote_code=trust)
         else:
-            ds = load_dataset(ds_spec)
+            trust = "/" in ds_spec
+            ds = load_dataset(ds_spec, trust_remote_code=trust)
     except Exception as e:
         # Fallback: if it looks like 'name:config' but was treated as a local path, try explicit two-arg call
         err_msg = str(e)
@@ -78,7 +81,8 @@ def _train_ner_lora(
             try:
                 ds_name, ds_config = [s.strip() for s in ds_spec.split(":", 1)]
                 log(f"Retrying with split name/config: {ds_name}, {ds_config}")
-                ds = load_dataset(ds_name, ds_config)
+                trust = "/" in ds_name
+                ds = load_dataset(ds_name, ds_config, trust_remote_code=trust)
             except Exception as e2:
                 log(f"Retry failed: {e2}")
                 raise