ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: auto-detect multimodal datasets in /check-format without requiring is_vlm flag

Browse source
This commit is contained in:
Roland Tannous 2026-02-13 17:29:39 +00:00
commit 4e4fc367b6
3 changed files with 17 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

2
studio/backend/models/datasets.py
View file

@ -18,6 +18,8 @@ class CheckFormatResponse(BaseModel):
requires_manual_mapping: bool
detected_format: str
columns: List[str]
is_multimodal: bool = False
multimodal_columns: Optional[List[str]] = None
suggested_mapping: Optional[Dict[str, str]] = None
detected_image_column: Optional[str] = None
detected_text_column: Optional[str] = None

2
studio/backend/routes/datasets.py
View file

@ -144,6 +144,8 @@ async def check_format(request: CheckFormatRequest):
requires_manual_mapping=result["requires_manual_mapping"],
detected_format=result["detected_format"],
columns=result["columns"],
is_multimodal=result.get("is_multimodal", False),
multimodal_columns=result.get("multimodal_columns"),
suggested_mapping=result.get("suggested_mapping"),
detected_image_column=result.get("detected_image_column"),
detected_text_column=result.get("detected_text_column"),

13
studio/backend/utils/datasets/dataset_utils.py
View file

@ -63,6 +63,11 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"""
columns = list(dataset.column_names) if hasattr(dataset, 'column_names') else list(next(iter(dataset)).keys())
# Auto-detect multimodal data regardless of is_vlm flag
multimodal_info = detect_multimodal_dataset(dataset)
if multimodal_info["is_multimodal"]:
is_vlm = True # Route to VLM detection automatically
if is_vlm:
vlm_structure = detect_vlm_dataset_structure(dataset)
requires_mapping = vlm_structure["format"] == "unknown"
@ -74,6 +79,8 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"suggested_mapping": None,
"detected_image_column": vlm_structure.get("image_column"),
"detected_text_column": vlm_structure.get("text_column"),
"is_multimodal": multimodal_info["is_multimodal"],
"multimodal_columns": multimodal_info.get("multimodal_columns"),
}
else:
# LLM flow
@ -91,6 +98,8 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"suggested_mapping": heuristic_mapping,
"detected_image_column": None,
"detected_text_column": None,
"is_multimodal": False,
"multimodal_columns": None,
}
else:
# Both detection and heuristic failed
@ -101,6 +110,8 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"suggested_mapping": None,
"detected_image_column": None,
"detected_text_column": None,
"is_multimodal": False,
"multimodal_columns": None,
}
# Known format detected
@ -111,6 +122,8 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"suggested_mapping": None,
"detected_image_column": None,
"detected_text_column": None,
"is_multimodal": False,
"multimodal_columns": None,
}
def _apply_user_mapping(dataset, mapping: dict, batch_size: int = 1000):