src/contrastive_capture.py

"""Regex-step activation capture — v8b, 32B-like mode.

This is closer to the QwQ-32B regex route:

  regex is NOT the token label itself.
  regex only locates reflection steps.
  Positive tokens are sampled from the whole regex-hit step/segment.
  Negative tokens are sampled from clean non-reflection steps/segments.

Compared with marker-token mode:
  old 8B: positive = regex marker token +/- window, negative = non-marker token, 1:1
  this:   positive = reflection-step all-token sample, negative = non-reflection-step all-token sample, default 1:2

The function signature is unchanged, so scripts/01/02/03 do not need structural edits.
"""
from typing import Dict, List, Tuple
import os
import re
import torch
from tqdm import tqdm


def _sample_even(xs: List[int], n: int) -> List[int]:
    if n <= 0 or not xs:
        return []
    if len(xs) <= n:
        return list(xs)
    step = len(xs) / float(n)
    return [xs[min(len(xs) - 1, int(i * step))] for i in range(n)]


def _merge_ranges(ranges):
    if not ranges:
        return []
    ranges = sorted((int(a), int(b)) for a, b in ranges if int(b) > int(a))
    merged = []
    for a, b in ranges:
        if not merged or a > merged[-1][1]:
            merged.append([a, b])
        else:
            merged[-1][1] = max(merged[-1][1], b)
    return [(a, b) for a, b in merged]


def _overlap(a, b, c, d):
    return a < d and c < b


def _in_any_range(a, b, ranges):
    for c, d in ranges:
        if _overlap(a, b, c, d):
            return True
    return False


def _expand_hit_to_step(text: str, start: int, end: int,
                        left_chars: int = 420,
                        right_chars: int = 780):
    """Expand regex hit to a local reasoning step.

    Boundary priority:
      1. nearby newline / paragraph boundary
      2. nearby sentence punctuation
      3. char-window fallback

    This avoids capturing only 'wait'/'maybe' tokens, while also avoiding
    whole-CoT capture that would OOM.
    """
    n = len(text)
    start = max(0, min(start, n))
    end = max(start, min(end, n))

    left_floor = max(0, start - left_chars)
    right_ceil = min(n, end + right_chars)

    # left boundary: last newline or sentence-like delimiter before hit
    left_candidates = [left_floor]
    for pat in [r"\n\s*\n", r"\n", r"(?<=[\.\!\?。！？])\s+"]:
        last = None
        for m in re.finditer(pat, text[left_floor:start]):
            last = m
        if last is not None:
            left_candidates.append(left_floor + last.end())
    seg_start = max(left_candidates)

    # right boundary: first newline or sentence-like delimiter after hit
    right_candidates = [right_ceil]
    suffix = text[end:right_ceil]
    for pat in [r"\n\s*\n", r"\n", r"(?<=[\.\!\?。！？])\s+"]:
        m = re.search(pat, suffix)
        if m is not None:
            right_candidates.append(end + m.end())
    seg_end = min(right_candidates)

    if seg_end <= seg_start:
        seg_start, seg_end = left_floor, right_ceil
    return seg_start, seg_end


def _basic_segments(text: str, max_len: int = 900):
    """Return coarse line/sentence segments for negative sampling."""
    segs = []
    n = len(text)
    i = 0
    while i < n:
        # skip whitespace
        while i < n and text[i].isspace():
            i += 1
        if i >= n:
            break

        # prefer newline as step boundary
        j_new = text.find("\n", i)
        if j_new == -1:
            j_new = n

        chunk_start, chunk_end = i, j_new
        # split overly long lines by sentence-ish punctuation
        if chunk_end - chunk_start <= max_len:
            segs.append((chunk_start, chunk_end))
        else:
            k = chunk_start
            while k < chunk_end:
                sub_end = min(chunk_end, k + max_len)
                window = text[k:sub_end]
                # try to stop at punctuation inside the window
                cut = None
                for m in re.finditer(r"[\.\!\?。！？]\s+", window):
                    cut = m.end()
                if cut is not None and k + cut > k + 40:
                    sub_end = k + cut
                segs.append((k, sub_end))
                k = sub_end

        i = j_new + 1

    return [(a, b) for a, b in segs if b > a]


def _token_positions_by_ranges(offsets, ranges, skip_head: int):
    pos = []
    for i, off in enumerate(offsets):
        if i < skip_head:
            continue
        if off is None:
            continue
        try:
            a, b = int(off[0]), int(off[1])
        except Exception:
            continue
        if b <= a:
            continue
        if _in_any_range(a, b, ranges):
            pos.append(i)
    return pos


def collect_contrastive_activations(
    model, tokenizer, pairs: List[dict], layers: List[int], device: str,
    samples_per_cot: int = 64, max_seq_len: int = 4096,
    skip_head: int = 16, logger=None,
) -> Tuple[Dict[int, Dict[str, torch.Tensor]], Dict[str, int]]:
    from configs import get_config
    from src.detectors import BehaviorDetector
    from src.utils import think_segment

    cfg = get_config("monitoring")
    detector = BehaviorDetector(cfg)

    # 32B-like defaults.
    # POS_PER_COT controls reflection-step all-token sample.
    # NEG_MULTIPLIER=2 means non-reflection tokens are twice positives.
    pos_per_cot = int(os.environ.get("REGEX_POS_PER_COT", str(samples_per_cot)))
    neg_multiplier = float(os.environ.get("REGEX_NEG_MULTIPLIER", "2.0"))
    step_left = int(os.environ.get("REGEX_STEP_LEFT_CHARS", "420"))
    step_right = int(os.environ.get("REGEX_STEP_RIGHT_CHARS", "780"))
    max_segment_len = int(os.environ.get("REGEX_MAX_SEGMENT_CHARS", "900"))

    model.eval()
    per_layer_acts = {L: [] for L in layers}
    per_layer_labels = {L: [] for L in layers}

    total_pos = 0
    total_neg = 0
    total_pos_all = 0
    total_neg_all = 0
    texts_seen = 0
    texts_used = 0
    skipped_no_regex = 0
    skipped_no_negative = 0

    if logger:
        logger.info("  regex capture mode = 32B-like step/segment all-token")
        logger.info(f"  REGEX_POS_PER_COT = {pos_per_cot}")
        logger.info(f"  REGEX_NEG_MULTIPLIER = {neg_multiplier}")
        logger.info(f"  step expansion chars = left {step_left}, right {step_right}")
        logger.info(f"  max negative segment chars = {max_segment_len}")

    def _forward_one(text: str):
        nonlocal total_pos, total_neg, total_pos_all, total_neg_all
        nonlocal texts_seen, texts_used, skipped_no_regex, skipped_no_negative

        text = think_segment(text or "")
        if not text:
            return
        texts_seen += 1

        det = detector.detect(text)
        spans = det.get("spans", [])
        if not spans:
            skipped_no_regex += 1
            return

        # Positive ranges: whole local step around each regex hit.
        pos_ranges = []
        for sp in spans:
            s, e = int(sp.get("start", -1)), int(sp.get("end", -1))
            if s >= 0 and e > s:
                pos_ranges.append(
                    _expand_hit_to_step(
                        text, s, e,
                        left_chars=step_left,
                        right_chars=step_right,
                    )
                )
        pos_ranges = _merge_ranges(pos_ranges)
        if not pos_ranges:
            skipped_no_regex += 1
            return

        # Negative ranges: clean line/sentence segments that do not overlap reflection ranges.
        neg_ranges = []
        for a, b in _basic_segments(text, max_len=max_segment_len):
            if not _in_any_range(a, b, pos_ranges):
                neg_ranges.append((a, b))
        neg_ranges = _merge_ranges(neg_ranges)

        try:
            enc = tokenizer(
                text,
                return_tensors=None,
                add_special_tokens=False,
                truncation=True,
                max_length=max_seq_len,
                return_offsets_mapping=True,
            )
            ids = enc["input_ids"]
            offsets = enc.get("offset_mapping")
        except TypeError:
            # Fast tokenizer should support offsets. If not, skip rather than falling
            # back to marker-level labels.
            return

        if offsets is None or len(ids) < skip_head + 4:
            return

        pos_all = _token_positions_by_ranges(offsets, pos_ranges, skip_head)
        neg_all = _token_positions_by_ranges(offsets, neg_ranges, skip_head)

        if not pos_all:
            skipped_no_regex += 1
            return
        if not neg_all:
            skipped_no_negative += 1
            return

        pos = _sample_even(pos_all, pos_per_cot)
        neg_cap = max(1, int(round(len(pos) * neg_multiplier)))
        neg = _sample_even(neg_all, neg_cap)

        if not pos or not neg:
            return

        positions = pos + neg
        labels = torch.tensor([1] * len(pos) + [0] * len(neg), dtype=torch.long)

        input_ids = torch.tensor([ids], device=device)
        with torch.no_grad():
            outputs = model(input_ids, output_hidden_states=True)

        for L in layers:
            if L + 1 >= len(outputs.hidden_states):
                continue
            hs = outputs.hidden_states[L + 1][0].float().cpu()
            per_layer_acts[L].append(hs[positions])
            per_layer_labels[L].append(labels)

        total_pos += len(pos)
        total_neg += len(neg)
        total_pos_all += len(pos_all)
        total_neg_all += len(neg_all)
        texts_used += 1

    for pair in tqdm(pairs, desc=" Regex-step capture"):
        # Do not use high/low prompt labels. Both are sources of natural CoT text.
        _forward_one(pair.get("high_reflection_cot") or pair.get("cot") or pair.get("text") or "")
        _forward_one(pair.get("low_reflection_cot") or "")

    out = {}
    for L in layers:
        if not per_layer_acts[L]:
            continue
        out[L] = {
            "acts": torch.cat(per_layer_acts[L], dim=0),
            "labels": torch.cat(per_layer_labels[L], dim=0),
        }
        if logger:
            labels = out[L]["labels"]
            logger.info(
                f"  L{L:>2}: captured {labels.numel()} step tokens "
                f"(+:{int((labels == 1).sum())} -:{int((labels == 0).sum())})"
            )

    return out, {
        "pos": int(total_pos),
        "neg": int(total_neg),
        "pos_all_before_sampling": int(total_pos_all),
        "neg_all_before_sampling": int(total_neg_all),
        "texts_seen": int(texts_seen),
        "texts_used": int(texts_used),
        "texts_skipped_no_regex": int(skipped_no_regex),
        "texts_skipped_no_negative": int(skipped_no_negative),
        "label_source": "32b_like_regex_step_segment_all_token_sampling",
        "pos_per_cot": int(pos_per_cot),
        "neg_multiplier": float(neg_multiplier),
        "step_left_chars": int(step_left),
        "step_right_chars": int(step_right),
    }