feat(dataset): use Parquet hf:// for BigBio (bc5cdr, ncbi_disease) and add dataset dropdown

app.py

@@ -57,14 +57,20 @@ def _train_ner_lora(
     try:
         # Medical aliases -> BigBio NER configs
         alias_map = {
-            "bc5cdr": ("bigbio/bc5cdr", "bigbio_ner"),
-            "ncbi_disease": ("bigbio/ncbi_disease", "bigbio_ner"),
+            # Use Parquet conversion branch via hf:// scheme
+            "bc5cdr": "hf://datasets/bigbio/bc5cdr@refs/convert/parquet/bigbio_ner",
+            "ncbi_disease": "hf://datasets/bigbio/ncbi_disease@refs/convert/parquet/bigbio_ner",
         }
         lower_spec = ds_spec.lower()
         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, trust_remote_code=True)
+            base = alias_map[lower_spec]
+            log(f"Using alias mapping (parquet): {ds_spec} -> {base}")
+            data_files = {
+                "train": f"{base}/train-*.parquet",
+                "validation": f"{base}/validation-*.parquet",
+                "test": f"{base}/test-*.parquet",
+            }
+            ds = load_dataset("parquet", data_files=data_files)
         elif ":" in ds_spec:
             ds_name, ds_config = [s.strip() for s in ds_spec.split(":", 1)]
             # If loading from community repo, allow remote code
@@ -324,7 +330,17 @@ def build_ui():
         )
         with gr.Row():
             base_model = gr.Textbox(value=DEFAULT_BASE_MODEL, label="Base model")
-            dataset_name = gr.Textbox(value=DEFAULT_DATASET, label="Dataset (token classification)")
+            dataset_name = gr.Dropdown(
+                choices=[
+                    "bc5cdr",
+                    "ncbi_disease",
+                    "wikiann:en",
+                    "conll2003",
+                ],
+                value=DEFAULT_DATASET,
+                allow_custom_value=True,
+                label="Dataset (token classification)",
+            )
         with gr.Row():
             epochs = gr.Slider(minimum=1, maximum=3, step=1, value=1, label="Epochs")
             batch = gr.Slider(minimum=4, maximum=16, step=2, value=8, label="Batch size")