File size: 8,515 Bytes
d8ae160 b279884 d8ae160 b279884 d8ae160 b279884 d8ae160 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 | """Per-message probe overlays that paint assistant text with class colors.
Mirrors the integration shape of ``utils.contrast``: one overlay per assistant
message, attached as ``message["_probe_overlay"]`` and rendered inline by
``render_chat_message``. Overlays cover only the message body — special tokens
(role markers, BOS/EOS) are filtered out at build time.
"""
from __future__ import annotations
from dataclasses import dataclass
from html import escape
import torch
from utils.probe_trace import ConversationTrace
_CLASS_COLORS: tuple[tuple[int, int, int], ...] = (
(210, 60, 60),
(50, 110, 210),
(60, 170, 90),
(210, 150, 50),
(170, 80, 200),
(200, 80, 130),
(90, 180, 200),
(170, 170, 70),
)
_MAX_ALPHA = 0.55
_PROBE_CSS = (
"<style>"
".probe-tok{position:relative;border-radius:2px;padding:0 1px;"
"cursor:default;white-space:pre;}"
".probe-tok>.probe-tip{display:none;position:absolute;bottom:100%;"
"left:50%;transform:translateX(-50%);margin-bottom:4px;padding:2px 6px;"
"border-radius:3px;background:#222;color:#eee;font-size:0.72em;"
"font-family:ui-monospace,monospace;white-space:nowrap;pointer-events:none;"
"z-index:10;box-shadow:0 2px 6px rgba(0,0,0,0.3);}"
".probe-tok:hover>.probe-tip{display:block;}"
".probe-wrap{line-height:1.75;white-space:pre-wrap;word-break:break-word;}"
"</style>"
)
@dataclass(frozen=True)
class ProbeOverlay:
tokens: list[str]
labels: list[str | None]
is_regression: bool
attribute_name: str | None
# Classification fields (empty when is_regression).
probs: list[list[float]]
predicted: list[int]
binary: bool
# Regression field (empty when not is_regression).
values: list[float]
# ---------------------------------------------------------------------------
# Building overlays from a trace
# ---------------------------------------------------------------------------
def _body_indices(trace: ConversationTrace, start: int, end: int) -> list[int]:
"""Indices inside an assistant span, with special tokens dropped."""
return [i for i in range(start, end) if not bool(trace.is_special[i].item())]
def build_classification_overlays(
*,
trace: ConversationTrace,
probs: torch.Tensor,
predicted: torch.Tensor,
labels: list[str | None],
binary: bool,
attribute_name: str | None = None,
) -> list[ProbeOverlay]:
overlays: list[ProbeOverlay] = []
for start, end in trace.assistant_spans:
idx = _body_indices(trace, start, end)
if not idx:
continue
overlays.append(
ProbeOverlay(
tokens=[trace.tokens[i] for i in idx],
labels=list(labels),
is_regression=False,
attribute_name=attribute_name,
probs=[probs[i].tolist() for i in idx],
predicted=[int(predicted[i].item()) for i in idx],
binary=binary,
values=[],
)
)
return overlays
def build_regression_overlays(
*,
trace: ConversationTrace,
values: torch.Tensor,
labels: list[str | None],
attribute_name: str | None = None,
) -> list[ProbeOverlay]:
if values.ndim == 2 and values.shape[1] >= 1:
values = values[:, 0]
overlays: list[ProbeOverlay] = []
for start, end in trace.assistant_spans:
idx = _body_indices(trace, start, end)
if not idx:
continue
overlays.append(
ProbeOverlay(
tokens=[trace.tokens[i] for i in idx],
labels=list(labels),
is_regression=True,
attribute_name=attribute_name,
probs=[],
predicted=[],
binary=False,
values=[float(values[i].item()) for i in idx],
)
)
return overlays
def attach_overlays(messages: list[dict], overlays: list[ProbeOverlay]) -> None:
"""Attach one overlay to each assistant message, in order.
Requires a 1:1 match. If the counts don't line up (e.g. the chat template
doesn't mark assistant tokens), clear overlays so the caller can show a
clear status instead of painting the wrong message.
"""
assistant_idxs = [i for i, m in enumerate(messages) if m.get("role") == "assistant"]
clear_overlays(messages)
if not assistant_idxs or len(overlays) != len(assistant_idxs):
return
for msg_idx, overlay in zip(assistant_idxs, overlays, strict=True):
messages[msg_idx]["_probe_overlay"] = overlay
def clear_overlays(messages: list[dict]) -> None:
for message in messages:
message.pop("_probe_overlay", None)
# ---------------------------------------------------------------------------
# Rendering
# ---------------------------------------------------------------------------
def _label_for(labels: list[str | None], idx: int) -> str:
if 0 <= idx < len(labels) and labels[idx]:
return labels[idx]
return str(idx)
def _display_token(token: str) -> str:
return token.replace("Ġ", " ").replace("▁", " ")
def _background(
probs_row: list[float], pred_idx: int, *, binary: bool, num_classes: int
) -> str:
if binary:
score = probs_row[0] if len(probs_row) == 1 else probs_row[-1]
signed = score - 0.5
alpha = min(1.0, abs(signed) * 2) * _MAX_ALPHA
r, g, b = (210, 60, 60) if signed > 0 else (50, 110, 210)
else:
baseline = 1.0 / max(num_classes, 2)
confidence = probs_row[pred_idx] if 0 <= pred_idx < len(probs_row) else 0.0
normalized = max(0.0, (confidence - baseline) / max(1e-6, 1.0 - baseline))
alpha = normalized * _MAX_ALPHA
r, g, b = _CLASS_COLORS[pred_idx % len(_CLASS_COLORS)]
if alpha < 0.02:
return "transparent"
return f"rgba({r},{g},{b},{alpha:.3f})"
def _tooltip(probs_row: list[float], labels: list[str | None]) -> str:
if len(probs_row) == 1:
positive = probs_row[0]
positive_label = _label_for(labels, 0)
# Single-output sigmoid: synthesize the complementary class so the
# hover shows both label probabilities, not just one.
return escape(
f"{positive_label} {positive:.2f} · not {positive_label} {1 - positive:.2f}"
)
ranked = sorted(enumerate(probs_row), key=lambda item: item[1], reverse=True)
parts = [f"{_label_for(labels, idx)} {prob:.2f}" for idx, prob in ranked]
return escape(" · ".join(parts))
def _regression_background(value: float, normalizer: float) -> str:
"""Red for positive, blue for negative, alpha by |value| relative to span max."""
if normalizer <= 1e-9:
return "transparent"
intensity = min(1.0, abs(value) / normalizer) * _MAX_ALPHA
if intensity < 0.02:
return "transparent"
r, g, b = (210, 60, 60) if value >= 0 else (50, 110, 210)
return f"rgba({r},{g},{b},{intensity:.3f})"
def render_probe_html(overlay: ProbeOverlay) -> str:
"""Render the assistant message as colored token spans with hover tips."""
spans: list[str] = []
if overlay.is_regression:
normalizer = max((abs(v) for v in overlay.values), default=0.0)
attribute = overlay.attribute_name or (
overlay.labels[0] if overlay.labels and overlay.labels[0] else "prediction"
)
for token, value in zip(overlay.tokens, overlay.values, strict=True):
bg = _regression_background(value, normalizer)
tip = escape(f"{attribute}: {value:.3f}")
text = escape(_display_token(token))
spans.append(
f'<span class="probe-tok" style="background:{bg};">'
f'{text}<span class="probe-tip">{tip}</span></span>'
)
else:
num_classes = max(1, len(overlay.probs[0]) if overlay.probs else 1)
for token, probs_row, pred_idx in zip(
overlay.tokens, overlay.probs, overlay.predicted, strict=True
):
bg = _background(
probs_row, pred_idx, binary=overlay.binary, num_classes=num_classes
)
tip = _tooltip(probs_row, overlay.labels)
text = escape(_display_token(token))
spans.append(
f'<span class="probe-tok" style="background:{bg};">'
f'{text}<span class="probe-tip">{tip}</span></span>'
)
return _PROBE_CSS + '<div class="probe-wrap">' + "".join(spans) + "</div>"
|