Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: send blank seed cells to Data Recipe prompts as empty text (#12789)

Nilay15871914d8f4
diff --git a/studio/backend/core/data_recipe/service.py b/studio/backend/core/data_recipe/service.py
index 5b08e3c778b..b4ae256ead3 100644
--- a/studio/backend/core/data_recipe/service.py
+++ b/studio/backend/core/data_recipe/service.py
@@ -5,7 +5,9 @@ from __future__ import annotations
from core.training.account_jobs import account_path, managed_account, validate_recipe_access
import base64
+import functools
import io
+import math
import os
from pathlib import Path
from typing import Any
@@ -145,6 +147,77 @@ def _apply_data_designer_image_context_patch() -> None:
_IMAGE_CONTEXT_PATCHED = True
+def _blank_missing_prompt_value(value: Any) -> Any:
+ if value is None or (isinstance(value, float) and math.isnan(value)):
+ return ""
+ return value
+
+
+def _allow_empty_prompt(_rendered_text: str) -> None:
+ return None
+
+
+# Filters that stringify their input before finalize runs, so a missing cell would still render "None" / "nan".
+_STRINGIFYING_FILTERS = (
+ "capitalize",
+ "escape",
+ "forceescape",
+ "lower",
+ "replace",
+ "string",
+ "title",
+ "trim",
+ "truncate",
+ "urlencode",
+)
+
+
+def _blank_missing_filter_input(filter_func: Any) -> Any:
+ value_index = 1 if hasattr(filter_func, "jinja_pass_arg") else 0
+
+ @functools.wraps(filter_func)
+ def wrapper(*args: Any, **kwargs: Any) -> Any:
+ if len(args) > value_index:
+ args = (
+ *args[:value_index],
+ _blank_missing_prompt_value(args[value_index]),
+ *args[value_index + 1 :],
+ )
+ return filter_func(*args, **kwargs)
+
+ return wrapper
+
+
+def _apply_data_designer_prompt_blank_patch() -> None:
+ try:
+ from data_designer.engine.column_generators.utils.prompt_renderer import ( # pyright: ignore[reportMissingImports]
+ RecordBasedPromptRenderer,
+ )
+ except ImportError:
+ return
+
+ if getattr(RecordBasedPromptRenderer, "_unsloth_prompt_blank_patch_applied", False):
+ return
+
+ original_prepare = RecordBasedPromptRenderer.prepare_jinja2_multi_template_renderer
+
+ def _patched_prepare(self: Any, template_name: str, *args: Any, **kwargs: Any) -> None:
+ # render() prepares on every record; only patch the env this call creates, else filter wrappers nest per row.
+ already_prepared = self._template_prepared_in_multi_template_renderer(template_name)
+ original_prepare(self, template_name, *args, **kwargs)
+ if already_prepared:
+ return
+ env = self._render_func_registry[template_name].func.__self__
+ env.finalize = _blank_missing_prompt_value
+ env._assert_rendered_text_not_empty = _allow_empty_prompt
+ for name in _STRINGIFYING_FILTERS:
+ if name in env.filters:
+ env.filters[name] = _blank_missing_filter_input(env.filters[name])
+
+ RecordBasedPromptRenderer.prepare_jinja2_multi_template_renderer = _patched_prepare
+ setattr(RecordBasedPromptRenderer, "_unsloth_prompt_blank_patch_applied", True)
+
+
def _require_public_provider_endpoint(endpoint: str) -> None:
"""The recipe engine dials providers itself, so a managed account's endpoint cannot use the pinned
transport: require HTTPS, which binds the peer to its certificate rather than to a DNS answer that
@@ -415,6 +488,7 @@ def create_data_designer(recipe: dict[str, Any], *, artifact_path: str | None =
validate_recipe_access(recipe)
account_path(artifact_path)
_apply_data_designer_image_context_patch()
+ _apply_data_designer_prompt_blank_patch()
from data_designer.interface.data_designer import DataDesigner # pyright: ignore[reportMissingImports]
if artifact_path is None:
diff --git a/studio/backend/tests/test_data_recipe_blank_seed_prompt.py b/studio/backend/tests/test_data_recipe_blank_seed_prompt.py
new file mode 100644
index 00000000000..5478f22a6e0
--- /dev/null
+++ b/studio/backend/tests/test_data_recipe_blank_seed_prompt.py
@@ -0,0 +1,168 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+import threading
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+
+import pytest
+
+
[email protected](name = "provider")
+def fixture_provider():
+ prompts: list[str] = []
+ systems: list[str] = []
+
+ class Handler(BaseHTTPRequestHandler):
+ def do_POST(self):
+ body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
+ for message in body["messages"]:
+ if message["role"] == "user":
+ prompts.extend(part["text"] for part in message["content"])
+ elif message["role"] == "system":
+ content = message["content"]
+ systems.append(content if isinstance(content, str) else content[0]["text"])
+ reply = json.dumps(
+ {
+ "id": "c",
+ "object": "chat.completion",
+ "created": 0,
+ "model": "x",
+ "choices": [
+ {
+ "index": 0,
+ "message": {"role": "assistant", "content": "ok"},
+ "finish_reason": "stop",
+ }
+ ],
+ "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+ }
+ ).encode()
+ self.send_response(200)
+ self.send_header("Content-Type", "application/json")
+ self.send_header("Content-Length", str(len(reply)))
+ self.end_headers()
+ self.wfile.write(reply)
+
+ def log_message(self, *args):
+ pass
+
+ server = ThreadingHTTPServer(("127.0.0.1", 0), Handler)
+ threading.Thread(target = server.serve_forever, daemon = True).start()
+ yield f"http://127.0.0.1:{server.server_address[1]}/v1", prompts, systems
+ server.shutdown()
+
+
+def test_blank_seed_cells_reach_the_model_as_empty_text(tmp_path: Path, provider) -> None:
+ pytest.importorskip("data_designer")
+ from core.data_recipe.service import build_config_builder, create_data_designer
+
+ endpoint, prompts, _ = provider
+ seed = tmp_path / "seed.csv"
+ seed.write_text("question,year,context\nWho?,,\nWhen?,2019,ctx\n", encoding = "utf-8")
+ recipe = {
+ "model_providers": [{"name": "p", "endpoint": endpoint, "provider_type": "openai"}],
+ "model_configs": [
+ {"alias": "m", "model": "x", "provider": "p", "inference_parameters": {}}
+ ],
+ "seed_config": {
+ "source": {"seed_type": "local", "path": str(seed)},
+ "sampling_strategy": "ordered",
+ },
+ "columns": [
+ {
+ "column_type": "llm-text",
+ "name": "answer",
+ "model_alias": "m",
+ "prompt": "{{ question }} ({{ year }}) {{ context }}",
+ }
+ ],
+ }
+
+ designer = create_data_designer(recipe, artifact_path = str(tmp_path / "artifacts"))
+ designer.preview(build_config_builder(recipe), num_records = 2)
+
+ assert "Who? () " in prompts
+ assert any(p.startswith("When? (2019") and p.endswith(") ctx") for p in prompts)
+ assert not [p for p in prompts if "None" in p or "nan" in p]
+
+
+def test_blank_system_prompt_cell_keeps_the_row(tmp_path: Path, provider) -> None:
+ pytest.importorskip("data_designer")
+ from core.data_recipe.service import build_config_builder, create_data_designer
+
+ endpoint, prompts, systems = provider
+ seed = tmp_path / "seed.csv"
+ seed.write_text("question,context\nWho?,\nWhen?,ctx\n", encoding = "utf-8")
+ recipe = {
+ "model_providers": [{"name": "p", "endpoint": endpoint, "provider_type": "openai"}],
+ "model_configs": [
+ {"alias": "m", "model": "x", "provider": "p", "inference_parameters": {}}
+ ],
+ "seed_config": {
+ "source": {"seed_type": "local", "path": str(seed)},
+ "sampling_strategy": "ordered",
+ },
+ "columns": [
+ {
+ "column_type": "llm-text",
+ "name": "answer",
+ "model_alias": "m",
+ "system_prompt": "{{ context }}",
+ "prompt": "{{ question }}",
+ }
+ ],
+ }
+
+ designer = create_data_designer(recipe, artifact_path = str(tmp_path / "artifacts"))
+ result = designer.preview(build_config_builder(recipe), num_records = 2)
+
+ assert sorted(result.dataset["question"]) == ["When?", "Who?"]
+ assert "Who?" in prompts and "When?" in prompts
+ assert "ctx" in systems
+ assert "None" not in systems
+
+
+ "template", ["{{ a | string }}", "{{ a | lower }}", "{{ a | title }}", "{{ a | trim }}"]
+)
+def test_stringifying_filters_render_blank_cells_as_empty_text(template: str) -> None:
+ pytest.importorskip("data_designer")
+ from core.data_recipe.service import _apply_data_designer_prompt_blank_patch
+ from data_designer.engine.column_generators.utils.prompt_renderer import (
+ PromptType,
+ RecordBasedPromptRenderer,
+ )
+ from data_designer.engine.models.recipes.response_recipes import TextResponseRecipe
+
+ _apply_data_designer_prompt_blank_patch()
+
+ def render(value):
+ return RecordBasedPromptRenderer(TextResponseRecipe()).render(
+ prompt_template = f"<{template}>", record = {"a": value}, prompt_type = PromptType.USER_PROMPT
+ )
+
+ assert render(None) == "<>"
+ assert render(float("nan")) == "<>"
+ assert render("Ab c") not in ("<>", "<None>")
+
+
+def test_reused_renderer_does_not_rewrap_filters_per_record() -> None:
+ pytest.importorskip("data_designer")
+ from core.data_recipe.service import _apply_data_designer_prompt_blank_patch
+ from data_designer.engine.column_generators.utils.prompt_renderer import (
+ PromptType,
+ RecordBasedPromptRenderer,
+ )
+ from data_designer.engine.models.recipes.response_recipes import TextResponseRecipe
+
+ _apply_data_designer_prompt_blank_patch()
+ renderer = RecordBasedPromptRenderer(TextResponseRecipe())
+ for _ in range(1500):
+ rendered = renderer.render(
+ prompt_template = "{{ a | lower }}",
+ record = {"a": None},
+ prompt_type = PromptType.USER_PROMPT,
+ )
+ assert rendered == ""