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

Studio: r16 fixes - inline fence tracking, broaden direct-intent plan with first/step prefixes

Browse source 
- _has_unclosed_code_fence() now scans every line with re.search and a
  shared FENCE_RUN regex, so an inline opening fence such as
  "First, let me write it. \`\`\`python" is tracked alongside the
  column-0 openers. A numbered list emitted INSIDE an inline-open
  fence no longer reads as a final answer.

- _DIRECT_NUMBERED_PLAN_FRAMING adds "first" and "step N(:?)" to its
  intent prefixes and "look up" to its verb whitelist. Plans like
  "First, analyze the uploaded CSV:\n1. Load rows\n2. Compute total"
  or "I'll look that up:\n1. Search the docs" now re-prompt instead
  of being mis-classified as final answers. The verb whitelist still
  excludes bare search/find/check/verify so "First, use binary
  search:\n1. Search the left half" stays an answer.
This commit is contained in:
Daniel Han 2026-05-24 20:47:34 +00:00
commit 7b0ed8333a
2 changed files with 76 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

19
studio/backend/core/inference/llama_cpp.py
View file

@ -180,9 +180,11 @@ _EXPLICIT_PLAN_HEADER = re.compile(
# poem" is not misclassified as a stall.
_DIRECT_NUMBERED_PLAN_FRAMING = re.compile(
r"\b(?:i['’](?:ll|m going to|m gonna)|i am (?:going to|gonna)|"
r"i will|i shall|let me|allow me|now i|next i)\b"
r"i will|i shall|let me|allow me|now i|next i|"
r"first|step \d+:?)\b"
r"[^\r\n]{0,160}"
r"\b(?:do (?:this|these|the following|it)|"
r"look (?:this |that |it |them )?up|"
r"proceed|start|begin|"
r"create|build|implement|set up|add|"
r"calculate|compute|analy[sz]e|parse|load|run|execute|test)\b"
@ -192,7 +194,9 @@ _DIRECT_NUMBERED_PLAN_FRAMING = re.compile(
)
_FENCE_LINE_RE = re.compile(r"^[ \t]*(?P<fence>`{3,}|~{3,})(?P<trailing>[^\r\n]*)$")
_FENCE_RUN_RE = re.compile(
r"(?<!`)(?P<backticks>`{3,})(?!`)|(?<!~)(?P<tildes>~{3,})(?!~)"
)
def _has_unclosed_code_fence(text: str) -> bool:
@ -201,16 +205,19 @@ def _has_unclosed_code_fence(text: str) -> bool:
A complete fence answer is already caught by _HAS_ANSWER_ARTIFACT.
This helper exists so that an OPEN fence (model still streaming
code, or stream cut short) does not let an embedded numbered list
inside the fence body masquerade as a final answer.
inside the fence body masquerade as a final answer. The scan is
done per line and uses ``search`` (not ``match``) so an inline
opener such as ``First, let me write it. \\`\\`\\`python`` is also
tracked.
"""
active_char: Optional[str] = None
active_len = 0
for line in text.splitlines():
m = _FENCE_LINE_RE.match(line)
m = _FENCE_RUN_RE.search(line)
if not m:
continue
fence = m.group("fence")
trailing = m.group("trailing").strip()
fence = m.group("backticks") or m.group("tildes")
trailing = line[m.end():].strip()
ch = fence[0]
if active_char is None:
active_char = ch