scripts/dedup.py

"""Shared deterministic dedup + precision-reject helpers.

Used by both the one-shot pool merger and the production corruption pipeline so
the two can never drift apart on what counts as a duplicate or a forbidden span.
"""
import re
from collections import defaultdict
from typing import List

# --- Tier-1 parity (vox app/src-tauri/src/services/stt/itn.rs) -----------------
# INVARIANT: these two regexes must stay byte-equivalent to itn.rs, otherwise the
# dataset gate will pass spans that production has already stripped.
NON_LEXICAL_FILLER_RE = re.compile(r"(?i)\b(um+|uh+|er+|ah+|hmm+)\b[,;]?\s*")
PUNCT_TRIM = ".,;:!?\"'()"

# Tier 1 (vox itn.rs) already strips these deterministically — a synthetic span
# matching them is a wasted training signal, so it is rejected before the judge.
TIER1 = re.compile(r"^[,.\s]*(um+|uh+|er+|ah+|hmm+)[,.\s]*$", re.IGNORECASE)
# Words whose deletion is never safe: pronouns, articles, and the conjunctions that
# invariant 7 protects. Deleting a bare "the," or "and," is far more likely to be a
# wrong label than a real disfluency, so such examples are dropped, not repaired.
PROTECTED = {"it", "they", "them", "he", "she", "we", "you", "a", "an", "the",
             "and", "but", "so", "i", "my", "your", "his", "her", "our", "their",
             "me", "him", "us", "them"}
PRON = {"it", "they", "them", "he", "she", "we", "you", "a", "an", "the"}
NONWORD_RE = re.compile(r"[^a-z0-9 ]+")
# Any surviving Tier-1 hesitation marker. Tier 1 deletes these upstream, so a label
# containing one describes text that can never reach Tier 2 in production.
TIER1_ANY = re.compile(r"\b(um+|uh+|er+|ah+|hmm+)\b", re.IGNORECASE)


def norm_tokens(text: str) -> frozenset:
    """Normalization for near-dedup: lowercase, strip punctuation, token set."""
    return frozenset(NONWORD_RE.sub(" ", text.lower()).split())


def near_dup(cand: frozenset, kept: List[frozenset], threshold: float = 0.85) -> bool:
    """Token-set Jaccard against kept fingerprints.

    Exact-text dedup is insufficient: template collapse emits near-identical
    sentences ("Deploy on Tuesday wait no Wednesday." x6 variants).
    """
    if not cand:
        return True
    for other in kept:
        inter = len(cand & other)
        if inter and inter / len(cand | other) >= threshold:
            return True
    return False


CUE_RE = re.compile(
    r"(actually|wait no|i mean|sorry|scratch that|rather|correction|"
    r"no,|no wait|^no—|—.*—)", re.IGNORECASE)
EMDASH_RE = re.compile(r"[—–-]")
FILLER_ONLY = re.compile(r"^[,.\s]*(like|so|well|you know)[,.\s]*$", re.IGNORECASE)


def shape_reject(ex) -> str:
    """Pre-judge structural screen: spans must LOOK like disfluencies.

    Bare grammar fixes ("make"->"makes"), synonym swaps ("middle"->"centre")
    and insertions with no disfluency marker die here mechanically instead of
    depending on judge consistency.
    """
    for s in ex.spans:
        t = s.text.strip()
        if ex.category == "REPAIR":
            if not CUE_RE.search(ex.raw_text):
                return f"REPAIR without audible cue in raw: {t[:40]!r}"
            if not CUE_RE.search(t) and not EMDASH_RE.search(t):
                return f"REPAIR span carries no cue: {t[:40]!r}"
            if not (s.target or "").strip():
                return "REPAIR span missing target"
        else:  # DELETE: filler-only, cue-led supersession, or truncated false start
            if FILLER_ONLY.match(t):
                continue
            if CUE_RE.search(t):
                continue
            if EMDASH_RE.search(t):
                continue
            return f"DELETE span has no disfluency shape: {t[:40]!r}"
    return ""


def apply_tier1_prestrip(text: str) -> str:
    """Simulate the two Tier-1 steps that run BEFORE Tier 2, without ITN.

    Mirrors vox `itn.rs`: strip non-lexical hesitation markers (um/uh/er/ah/hmm),
    then collapse consecutive identical word tokens ("at at" -> "at").

    Purpose: a Tier-2 span that this function already deletes can never appear in
    production, so training on it is wasted signal that also inflates eval scores
    with a task the model will never be asked to do.
    """
    text = NON_LEXICAL_FILLER_RE.sub("", text)
    out: List[str] = []
    words = text.split()
    i = 0
    while i < len(words):
        cur = words[i].strip(PUNCT_TRIM)
        if i + 1 < len(words) and cur and cur.lower() == words[i + 1].strip(PUNCT_TRIM).lower():
            i += 1
            continue
        out.append(words[i])
        i += 1
    return " ".join(out)


def already_gone_by_tier1(raw_text: str, span_text: str) -> bool:
    """True if Tier 1 would have removed this span before Tier 2 ever sees it."""
    stripped = span_text.strip()
    if not stripped:
        return False
    return stripped not in apply_tier1_prestrip(raw_text)


def deterministic_reject(ex) -> str:
    """Precision-first pre-judge screen. Returns reason ('' = pass)."""
    if TIER1_ANY.search(ex.clean_text or ""):
        return "Tier-1 hesitation marker survives in clean_text"
    for s in ex.spans:
        t = s.text.strip()
        if TIER1.match(t) or TIER1_ANY.search(t):
            return f"tier1-duplicative span {t!r}"
        if t.lower().strip(",. ") in PROTECTED:
            return f"protected-word span (pronoun/article/conjunction) {t!r}"
    return ""


class NearDup:
    """Exact Jaccard >= t rejection, bucketed by token count.

    J(A,B) <= min(|A|,|B|) / max(|A|,|B|), so any pair reaching J >= t must have
    token counts within a factor t. Bucketing on length is therefore lossless and
    turns the O(n^2) sweep into per-bucket sweeps, which matters at Layer 3 scale.
    """

    def __init__(self, threshold: float = 0.85):
        self.t = threshold
        self.buckets: dict[int, list] = defaultdict(list)

    def seen(self, fp: frozenset) -> bool:
        n = len(fp)
        if not n:
            return True
        lo = max(1, int(n * self.t))
        for size in range(lo, int(n / self.t) + 2):
            for other in self.buckets.get(size, ()):
                inter = len(fp & other)
                if inter and inter / len(fp | other) >= self.t:
                    return True
        return False

    def add(self, fp: frozenset) -> None:
        self.buckets[len(fp)].append(fp)