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.

Normalize role names when converting ChatML to Alpaca (#14304)

ShryukGrandhie6bfccdd725b
diff --git a/studio/backend/tests/test_chatml_to_alpaca.py b/studio/backend/tests/test_chatml_to_alpaca.py
index 618ea950afd..8afb9dbb7c7 100644
--- a/studio/backend/tests/test_chatml_to_alpaca.py
+++ b/studio/backend/tests/test_chatml_to_alpaca.py
@@ -81,6 +81,26 @@ def test_a_single_exchange_converts_as_before():
]
+def test_role_names_are_matched_case_and_whitespace_insensitively():
+ dataset = Dataset.from_dict(
+ {
+ "messages": [
+ [
+ {"role": "System", "content": "Be brief."},
+ {"role": "User", "content": "Hello"},
+ {"role": " ASSISTANT ", "content": "Hi there"},
+ ]
+ ]
+ }
+ )
+
+ converted = convert_chatml_to_alpaca(dataset, num_proc = 1)
+
+ assert converted.to_list() == [
+ {"instruction": "Hello", "input": "Be brief.", "output": "Hi there"}
+ ]
+
+
def test_unpaired_empty_and_unknown_turns_are_skipped():
dataset = Dataset.from_dict(
{
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index d4a5afa2381..ec13b113644 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -241,7 +241,7 @@ def convert_chatml_to_alpaca(
for convo, column_system in zip(chatml_data, column_systems):
turns = []
for msg in convo or []:
- role = roles.get(msg.get("role") or msg.get("from"))
+ role = roles.get(_normalize_role_alias(msg.get("role") or msg.get("from")))
content = _content_text(msg.get("content") or msg.get("value"))
if role is None or not content:
continue