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: keep a system column when the advisor mapping carries metadata (#14999)

Vineeth Sai Varikuntlaecf0054bb5dc
diff --git a/studio/backend/tests/test_alpaca_ai_assist_mapping.py b/studio/backend/tests/test_alpaca_ai_assist_mapping.py
index dcab1c7e10d..ac6f6b1d29e 100644
--- a/studio/backend/tests/test_alpaca_ai_assist_mapping.py
+++ b/studio/backend/tests/test_alpaca_ai_assist_mapping.py
@@ -140,3 +140,49 @@ def test_chatml_advisor_mapping_renders_list_cells_as_text():
assert result["final_format"] == "chatml_conversations", result["warnings"]
assert [convo[-1]["content"] for convo in result["dataset"]["conversations"]] == ["0", "1, 2"]
+
+
+def test_chatml_advisor_mapping_keeps_a_system_column():
+ result = format_dataset(
+ Dataset.from_dict({"persona": ["Be a pirate.", ""], "text": ["hi", "yo"], "label": [1, 0]}),
+ format_type = "chatml",
+ batch_size = 2,
+ custom_format_mapping = {
+ "persona": "system",
+ "text": "user",
+ "label": "assistant",
+ "__label_mapping": LABEL_MAPPING,
+ },
+ )
+
+ assert result["final_format"] == "chatml_conversations", result["warnings"]
+ assert list(result["dataset"]["conversations"]) == [
+ [
+ {"role": "system", "content": "Be a pirate."},
+ {"role": "user", "content": "hi"},
+ {"role": "assistant", "content": "pos"},
+ ],
+ [
+ {"role": "user", "content": "yo"},
+ {"role": "assistant", "content": "neg"},
+ ],
+ ]
+
+
+def test_chatml_advisor_mapping_puts_the_system_prompt_before_a_system_column():
+ result = format_dataset(
+ Dataset.from_dict({"persona": ["Be a pirate."], "text": ["hi"], "label": [1]}),
+ format_type = "chatml",
+ batch_size = 1,
+ custom_format_mapping = {
+ "persona": "system",
+ "text": "user",
+ "label": "assistant",
+ "__system_prompt": PROMPT,
+ },
+ )
+
+ assert result["dataset"]["conversations"][0][0] == {
+ "role": "system",
+ "content": f"{PROMPT}\n\nBe a pirate.",
+ }
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index b853403c6fd..c75e35c534d 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -265,11 +265,11 @@ def _apply_template_mapping(
meta: dict,
batch_size: int = 1000,
):
- """Apply advisor-driven mapping for non-conversational datasets: groups columns by assigned role (user/assistant), concatenates values within each role into one message, and injects an optional system prompt. Label mapping converts integer labels to human-readable strings. Returns a dataset with a single 'conversations' column."""
+ """Apply advisor-driven mapping for non-conversational datasets: groups columns by assigned role (system/user/assistant), concatenates values within each role into one message, and injects an optional system prompt ahead of any system column. Label mapping converts integer labels to human-readable strings. Returns a dataset with a single 'conversations' column."""
system_prompt = meta.get("__system_prompt", "")
label_mapping = meta.get("__label_mapping", {}) # {col: {int_str: label_str}}
- role_groups: dict[str, list[str]] = {"user": [], "assistant": []}
+ role_groups: dict[str, list[str]] = {"system": [], "user": [], "assistant": []}
for col, role in column_roles.items():
canonical = _TO_CHATML.get(role, role)
if canonical in role_groups:
@@ -289,8 +289,17 @@ def _apply_template_mapping(
for i in range(num):
convo = []
- if system_prompt:
- convo.append({"role": "system", "content": system_prompt})
+ # The system prompt leads, then system columns, so neither is dropped.
+ system_parts = [
+ _extract_column_value(examples[col][i], col, label_mapping)
+ for col in role_groups["system"]
+ if col in examples
+ ]
+ system_text = "\n\n".join(
+ part for part in (system_prompt, "\n".join(p for p in system_parts if p)) if part
+ )
+ if system_text:
+ convo.append({"role": "system", "content": system_text})
user_parts = []
for col in role_groups["user"]: