fix: LLM-assisted mapping flows from /check-format to training

- Frontend auto-saves suggested_mapping into datasetManualMapping when
  check-format returns requires_manual_mapping=false, so the mapping
  flows to training via custom_format_mapping (no redundant AI calls)
- Backend returns meaningful warning when column detection fails
  (LLM-generated or static fallback) for both text and VLM datasets
- /check-format endpoint merges check_dataset_format warnings with
  existing URL-based image detection warnings

studio/backend/routes/datasets.py

@@ -438,18 +438,19 @@ def check_format(
         else:
             preview_samples = _serialize_preview_rows(preview_slice)
 
-        # Lightweight URL-based image detection for VLM datasets
-        warning = None
+        # Collect warnings: from check_dataset_format + URL-based image detection
+        warning = result.get("warning")
         image_col = result.get("detected_image_column")
         if image_col and image_col in (result.get("columns") or []):
             try:
                 sample_val = preview_slice[0][image_col]
                 if isinstance(sample_val, str) and sample_val.startswith(("http://", "https://")):
-                    warning = (
+                    url_warning = (
                         "This dataset contains image URLs instead of embedded images. "
                         "Images will be downloaded during training, which may be slow for large datasets."
                     )
                     logger.info(f"URL-based image column detected: {image_col}")
+                    warning = f"{warning} {url_warning}" if warning else url_warning
             except Exception:
                 pass
 

studio/backend/utils/datasets/dataset_utils.py

@@ -85,6 +85,21 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
         vlm_structure = detect_vlm_dataset_structure(dataset)
         requires_mapping = vlm_structure["format"] == "unknown"
 
+        warning = None
+        if requires_mapping:
+            img_col = vlm_structure.get("image_column")
+            txt_col = vlm_structure.get("text_column")
+            missing = []
+            if not img_col:
+                missing.append("image")
+            if not txt_col:
+                missing.append("text")
+            if missing:
+                warning = (
+                    f"Could not auto-detect {' or '.join(missing)} column. "
+                    "Please assign image and text columns manually."
+                )
+
         return {
             "requires_manual_mapping": requires_mapping,
             "detected_format": vlm_structure["format"],
@@ -94,6 +109,7 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
             "detected_text_column": vlm_structure.get("text_column"),
             "is_image": multimodal_info["is_image"],
             "multimodal_columns": multimodal_info.get("multimodal_columns"),
+            "warning": warning,
             **audio_fields,
         }
 
@@ -168,6 +184,31 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
                 import logging
                 logging.getLogger(__name__).debug(f"LLM column classification skipped: {e}")
 
+            # Both heuristics and LLM failed — generate a meaningful warning
+            warning = None
+            try:
+                from .llm_assist import llm_generate_dataset_warning
+                from itertools import islice
+                sample_rows = []
+                for s in islice(dataset, 3):
+                    row = {col: str(s[col])[:150] for col in s}
+                    sample_rows.append(row)
+                warning = llm_generate_dataset_warning(
+                    issues=[
+                        f"Could not auto-detect column roles from columns: {columns}",
+                        f"Sample values: {sample_rows[0] if sample_rows else 'N/A'}",
+                    ],
+                    modality="text",
+                    column_names=columns,
+                )
+            except Exception:
+                pass
+            if not warning:
+                warning = (
+                    f"Could not auto-detect column roles for columns: {columns}. "
+                    "Please assign roles (user, assistant, etc.) manually."
+                )
+
             return {
                 "requires_manual_mapping": True,
                 "detected_format": "unknown",
@@ -177,6 +218,7 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
                 "detected_text_column": None,
                 "is_image": False,
                 "multimodal_columns": None,
+                "warning": warning,
                 **audio_fields,
             }
 

studio/frontend/src/features/training/hooks/use-training-actions.ts

@@ -78,6 +78,19 @@ export function useTrainingActions() {
           });
         }
 
+        // Auto-save LLM/heuristic suggested mapping so training receives it
+        // as custom_format_mapping (avoids re-detecting during training).
+        if (!check.requires_manual_mapping && check.suggested_mapping) {
+          const hint: Record<string, string> = {};
+          const table = ROLE_REMAP[config.datasetFormat];
+          for (const [col, role] of Object.entries(check.suggested_mapping)) {
+            hint[col] = table ? (table[role] ?? role) : role;
+          }
+          if (Object.keys(hint).length > 0) {
+            useTrainingConfigStore.getState().setDatasetManualMapping(hint);
+          }
+        }
+
         if (check.requires_manual_mapping && !hasManualMapping(config, isVlm, isAudio)) {
           // Pre-fill from suggested_mapping or VLM detected columns
           const hint: Record<string, string> = {};