diff --git a/studio/backend/utils/datasets/llm_assist.py b/studio/backend/utils/datasets/llm_assist.py index 34ecd5eb55..ded4c49a00 100644 --- a/studio/backend/utils/datasets/llm_assist.py +++ b/studio/backend/utils/datasets/llm_assist.py @@ -562,7 +562,8 @@ def _run_multi_pass_advisor( {samples_text} Design a conversion strategy to turn this into conversation format for fine-tuning. - The strategy should create a system prompt, a user message template, and an assistant message template. + The strategy should create a user message template and an assistant message template. + Optionally include a system prompt ONLY if the task is ambiguous from the data alone. RULES: - Use {{column_name}} placeholders in templates to reference column values. @@ -574,10 +575,13 @@ def _run_multi_pass_advisor( - The assistant template should produce the expected model output. - column_roles: mark columns used in the user template as "user", columns used in the assistant template as "assistant". + - system_prompt: set to null if the user/assistant templates alone + make the task clear. Only provide one when extra context is needed + (e.g. persona, domain expertise, output format constraints). Respond with a JSON object: {{ - "system_prompt": "", + "system_prompt": "", "user_template": "