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

Studio: require newline after Plan: / Approach: header so inline product text does not re-prompt

Browse source 
After narrowing the colon marker to lines starting with a generic
determiner ("My plan:" / "The approach:" / ...), inline product or
pricing answers like "Your current Plan: Pro includes local chats",
"The plan: Basic is free, Pro is $10/month", or
"My plan: use dynamic programming" still slipped into the re-prompt
path and could wipe a valid answer.

Add a lookahead requiring a newline (with optional trailing horizontal
whitespace) after the colon, so only header-style framings like
"Plan:\n1. search\n2. summarise" or "My approach:\n1. fetch" count.
Inline "Plan: <text>" is now treated as ordinary prose.

Add eight regression samples (lesson plan, meal plan, marketing plan,
pricing plan, recommended approach, migration plan, dynamic-programming
plan, currently active plan) all of which previously re-prompted under
the unanchored matcher and now correctly do not.
This commit is contained in:
Daniel Han 2026-05-24 14:41:44 +00:00
commit 3dc26e7acf
2 changed files with 21 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

11
studio/backend/core/inference/llama_cpp.py
View file

@ -69,10 +69,11 @@ _INTENT_SIGNAL = re.compile(
r"here (?:is|are) (?:my |the |a )?(?:plan|approach|steps))"
r"|"
# Bare "Plan:" / "Approach:" (optionally preceded by a determiner
# like "My" / "The" / "Our") anchored to start of line so direct
# answers like "Here is a lesson plan:" or "meal plan:" do not trip
# the re-prompt path.
r"(?:^|\r?\n)[ \t]*(?:(?:my|the|our|a|this|that)\s+)?(?:plan|approach):"
# like "My" / "The" / "Our") anchored to start of line AND followed
# by a newline. Inline forms like "Your current Plan: Pro includes
# local chats" or "The plan: $10/month" must NOT trip the re-prompt
# path; only header-style framings ("Plan:\n1. ...") count.
r"(?:^|\r?\n)[ \t]*(?:(?:my|the|our|a|this|that)\s+)?(?:plan|approach):[ \t]*(?=\r?\n)"
r"|"
# "Now I" / "Next I" patterns
r"\b(?:now i|next i)\b"
@ -138,7 +139,7 @@ _PLAN_LIST_FRAMING = re.compile(
r"\b(?:search|look up|call|use|fetch|browse|run|execute|"
r"check|find|open|verify|compare|summari[sz]e|think|respond|"
r"answer|analy[sz]e|explore|outline|gather|query|reason)\b"
r"|(?:^|\r?\n)[ \t]*(?:(?:my|the|our|a|this|that)\s+)?(?:plan|approach):",
r"|(?:^|\r?\n)[ \t]*(?:(?:my|the|our|a|this|that)\s+)?(?:plan|approach):[ \t]*(?=\r?\n)",
re.IGNORECASE,
)

21
studio/backend/tests/test_llama_cpp_reprompt_guard.py
View file

@ -335,13 +335,14 @@ def test_reprompts_on_incomplete_html_intent():
def test_reprompts_on_plan_colon_intent():
"""Bare ``Plan:`` / ``Approach:`` at the start of a structured reply
is now an intent signal so the plan stall re-prompts. Pre-fix the
response slipped past ``_INTENT_SIGNAL`` entirely."""
"""Bare ``Plan:`` / ``Approach:`` followed by a newline at the start
of a structured reply is now an intent signal so the plan stall
re-prompts. Inline ``Plan: <text>`` (no newline) is NOT an intent
signal because that shape is common in marketing / product answers
such as ``Plan: Pro is $10/month``."""
samples = [
"Plan:\n1. search the docs\n2. summarise",
"Approach:\n1. fetch the data\n2. compare",
"Plan: search the docs then summarise",
]
for s in samples:
assert _INTENT_SIGNAL.search(s), s
@ -403,17 +404,25 @@ def test_plan_colon_intent_is_line_anchored():
re-prompt path and risk wiping a valid response."""
# These mid-line "plan:" / "approach:" mentions are NOT intent signals.
# The qualifier before "plan" is a content noun ("lesson", "meal",
# "migration") rather than a generic determiner ("my", "the", ...).
# "migration") rather than a generic determiner, OR the colon is
# followed by inline content instead of a newline-anchored header.
direct_answers = [
"Here is a lesson plan:\n1. Warm-up\n2. Group practice\n3. Assessment",
"I prepared a meal plan: rice, beans, eggs.",
"Quick approach: top-down then bottom-up.",
"Your current Plan: Pro includes local chats.",
"The plan: Basic is free, Pro is $10/month, Enterprise is custom.",
"My plan: use dynamic programming with memoisation.",
"Recommended approach: use the Python SDK for uploads.",
"The migration plan: backup, run, verify all in one window.",
]
for s in direct_answers:
assert not _INTENT_SIGNAL.search(s), s
assert not _would_reprompt(s), s
# "Plan:" / "Approach:" with optional generic determiner at the start
# of a line IS an intent signal.
# of a line, FOLLOWED BY A NEWLINE, IS an intent signal. The newline
# requirement is what filters inline product/answer text such as
# "The plan: Pro is $10/month" out of the intent path.
plan_starts = [
"Plan:\n1. search\n2. summarise",
"Approach:\n1. fetch\n2. compare",