File size: 14,033 Bytes
ac9bb45 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 | """
PIX-149: Gestalt Fusion Engine
Unifies the three core emotional intelligence models of Pixelated Empathy
into a single, real-time inference call:
1. PsyDefDetect — DMRS Defense Mechanism Classifier (DeBERTa)
2. Plutchik — 8-emotion wheel scoring (passed in externally)
3. OCEAN — Big Five personality trait scoring (passed in externally)
The fused ``GestaltState`` dataclass powers:
- PIX-147: WebSocket "Live X-Ray" resistance monitor
- PIX-148: Adversarial Persona Injection (dynamic defense-aware prompts)
- PIX-150: Empathy PQ metric validation (trainee scoring)
"""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional
import torch
from transformers import AutoTokenizer
from ai.training.defense_mechanisms.constants import DEFENSE_LABELS
from ai.training.defense_mechanisms.dataset import format_dialogue
from ai.training.defense_mechanisms.model import DefenseClassifier, DefensePrediction
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Domain enumerations
# ---------------------------------------------------------------------------
PLUTCHIK_EMOTIONS = frozenset(
{
"anger",
"anticipation",
"disgust",
"fear",
"joy",
"sadness",
"surprise",
"trust",
}
)
OCEAN_TRAITS = frozenset(
{
"openness",
"conscientiousness",
"extraversion",
"agreeableness",
"neuroticism",
}
)
class CrisisLevel(str, Enum):
"""
Behavioral risk level derived from the fused Gestalt state.
"""
NONE = "none"
ELEVATED = "elevated"
HIGH = "high"
ACUTE = "acute"
# ---------------------------------------------------------------------------
# Fused output dataclass
# ---------------------------------------------------------------------------
@dataclass
class GestaltState:
"""
Unified emotional-psychological state for a single dialogue turn.
"""
# --- Defense Mechanism (PsyDefDetect) ---
defense_label: int
defense_label_name: str
defense_confidence: float
defense_maturity: Optional[float]
defense_probabilities: dict[str, float]
# --- Emotion (Plutchik) ---
plutchik_scores: dict[str, float]
dominant_emotion: str
dominant_emotion_intensity: float
# --- Personality (OCEAN) ---
ocean_scores: dict[str, float]
# --- Fused behavioral outputs ---
crisis_level: CrisisLevel
behavioral_prediction: str
persona_directive: str
breakthrough_score: float
raw_metadata: dict = field(default_factory=dict, repr=False)
# ---------------------------------------------------------------------------
# Crisis & behavioral prediction logic
# ---------------------------------------------------------------------------
_ACTION_DEFENSE_LABEL = 1
_HIGH_ADAPTIVE_LABEL = 7
_DISAVOWAL_LABEL = 3
_MAJOR_IMAGE_DISTORTING_LABEL = 2
_CRISIS_AMPLIFYING_EMOTIONS = frozenset({"sadness", "fear", "anger", "disgust"})
_INJECTION_MATURITY_THRESHOLD = 0.43
_BREAKTHROUGH_MATURITY_THRESHOLD = 0.71
def _dominant_emotion(plutchik: dict[str, float]) -> tuple[str, float]:
"""Return the emotion with the highest intensity score."""
if not plutchik:
return "unknown", 0.0
dominant = max(plutchik, key=lambda k: plutchik[k])
return dominant, plutchik[dominant]
def _compute_crisis_level(
defense_label: int,
defense_maturity: Optional[float],
dominant_emotion: str,
dominant_intensity: float,
ocean_neuroticism: float,
) -> CrisisLevel:
"""Determine crisis level from the fused signals."""
is_action = defense_label == _ACTION_DEFENSE_LABEL
is_major_distorting = defense_label == _MAJOR_IMAGE_DISTORTING_LABEL
emotion_is_crisis = dominant_emotion in _CRISIS_AMPLIFYING_EMOTIONS
if is_action:
return (
CrisisLevel.ACUTE
if emotion_is_crisis
and dominant_intensity > 0.6
or ocean_neuroticism > 0.75
else CrisisLevel.HIGH
)
if is_major_distorting and emotion_is_crisis:
return CrisisLevel.HIGH
maturity = defense_maturity if defense_maturity is not None else 0.5
if maturity < 0.3 and emotion_is_crisis:
return CrisisLevel.ELEVATED
return CrisisLevel.NONE
def _behavioral_prediction(
defense_label_name: str,
dominant_emotion: str,
crisis_level: CrisisLevel,
defense_maturity: Optional[float],
) -> str:
"""Produce a concise human-readable behavioral prediction string."""
maturity_str = (
f"maturity={defense_maturity:.2f}" if defense_maturity is not None else "N/A"
)
base = f"{dominant_emotion.capitalize()} + {defense_label_name} ({maturity_str})"
if crisis_level == CrisisLevel.ACUTE:
return f"{base} → acute distress risk. Therapist must de-escalate immediately."
if crisis_level == CrisisLevel.HIGH:
return (
f"{base} → high-risk defensive pattern. "
"Therapist should slow pace and validate."
)
if crisis_level == CrisisLevel.ELEVATED:
return (
f"{base} → elevated defensiveness. "
"Therapist should use reflective listening."
)
if (
defense_maturity is not None
and defense_maturity >= _BREAKTHROUGH_MATURITY_THRESHOLD
):
return (
f"{base} → adaptive coping detected. "
"Therapist can introduced deeper reframing."
)
return f"{base} → neutral or intermediate defense. Continue with exploration."
def _persona_directive(
defense_label: int,
defense_label_name: str,
defense_maturity: Optional[float],
) -> str:
"""Return a system-prompt injection clause for PIX-148."""
if defense_maturity is None or defense_maturity >= _INJECTION_MATURITY_THRESHOLD:
return ""
directives: dict[int, str] = {
_ACTION_DEFENSE_LABEL: (
"[System: You are in an Action Defense state. "
"Act out, change subject. Do NOT acknowledge emotions directly.]"
),
_MAJOR_IMAGE_DISTORTING_LABEL: (
"[System: You are in a Major Image-Distorting state. "
"Split people into all-good or all-bad.]"
),
_DISAVOWAL_LABEL: (
"[System: You are in a Disavowal state. Deny emotional reality.]"
),
}
return directives.get(
defense_label,
f"[System: Maintain your '{defense_label_name}' defense mechanism.]",
)
def _breakthrough_score(
defense_maturity: Optional[float],
previous_maturity: Optional[float],
) -> float:
"""Score the magnitude of a positive maturity shift."""
if defense_maturity is None or previous_maturity is None:
return 0.0
delta = defense_maturity - previous_maturity
return max(0.0, min(1.0, delta))
def _validate_scores(
scores: dict[str, float],
expected_keys: frozenset[str],
label: str,
) -> dict[str, float]:
"""Validate normalized scores."""
for key, value in scores.items():
if not isinstance(value, (int, float)):
raise ValueError(f"{label}['{key}'] must be numeric")
if not 0.0 <= float(value) <= 1.0:
raise ValueError(f"{label}['{key}'] score {value} out of range [0, 1]")
if unknown := set(scores) - expected_keys:
logger.warning(f"{label} contains unknown keys {sorted(unknown)}")
return {k: float(v) for k, v in scores.items()}
# ---------------------------------------------------------------------------
# GestaltEngine
# ---------------------------------------------------------------------------
class GestaltEngine:
"""
Real-time Gestalt Fusion Engine for the Empathy Gym™.
"""
def __init__(self) -> None:
self._defense_model = None
self._defense_tokenizer = None
self._previous_maturity: Optional[float] = None
def load_defense_model(
self,
checkpoint_path: Optional[str] = None,
device: str = "cpu",
) -> None:
"""Load the PsyDefDetect model (optional if using NIM)."""
if not checkpoint_path:
logger.info("GestaltEngine: No checkpoint provided, using NIM by default.")
self._defense_model = DefenseClassifier()
return
try:
self._initialize_local_model(checkpoint_path, device)
except Exception as exc:
logger.error(
f"GestaltEngine: Failed to load checkpoint {checkpoint_path}: {exc}"
)
logger.info("GestaltEngine: Falling back to NIM-based DefenseClassifier.")
self._defense_model = DefenseClassifier()
def _initialize_local_model(self, checkpoint_path: str, device: str) -> None:
"""Initialize the legacy DeBERTa model from a local checkpoint."""
checkpoint = torch.load(
checkpoint_path, map_location=device, weights_only=False
)
config = checkpoint.get("config", {})
model_name = config.get("base_model", "microsoft/deberta-v3-base")
model = DefenseClassifier(
model_name=model_name,
num_labels=config.get("num_labels", 9),
r_drop_enabled=False,
)
# Handle PyTorch-based DefenseClassifier if checkpoint is provided
if hasattr(model, "load_state_dict"):
model.load_state_dict(checkpoint["model_state_dict"], strict=False)
model.to(device)
model.eval()
self._defense_tokenizer = AutoTokenizer.from_pretrained(model_name)
self._defense_model = model
logger.info(f"GestaltEngine: model loaded from {checkpoint_path}")
@property
def defense_model_loaded(self) -> bool:
return self._defense_model is not None
def reset_session(self) -> None:
self._previous_maturity = None
def _classify_defense(
self,
dialogue: list[dict[str, str]],
target_utterance: str,
max_turns: int = 40,
) -> tuple[int, str, float, Optional[float], dict[str, float]]:
"""Run PsyDefDetect inference."""
if self._defense_model is None:
raise RuntimeError(
"GestaltEngine: Defense model not initialized. "
"Call load_defense_model() first."
)
turns = [
{"speaker": t.get("speaker", "Unknown"), "text": t.get("text", "")}
for t in dialogue[-max_turns:]
]
# Ensure the target utterance is in the sequence for format_dialogue to mark it
target_normalized = target_utterance.strip().lower()
if all(t.get("text", "").strip().lower() != target_normalized for t in turns):
turns.append({"speaker": "User", "text": target_utterance})
formatted = format_dialogue(turns, target_utterance, max_turns)
# Handle the new text-based NIM classifier
if hasattr(self._defense_model, "nim"):
pred = self._defense_model.predict([formatted])[0]
else:
pred = self._legacy_inference(formatted)
prob_dict = {
DEFENSE_LABELS.get(i, str(i)): round(p, 4)
for i, p in enumerate(pred.probabilities)
}
return (
pred.label,
pred.label_name,
pred.confidence,
pred.maturity_score,
prob_dict,
)
def _legacy_inference(self, formatted_text: str) -> DefensePrediction:
"""Run inference using the local PyTorch DeBERTa model."""
if self._defense_tokenizer is None:
raise RuntimeError("GestaltEngine: No tokenizer for PyTorch model.")
encoding = self._defense_tokenizer(
formatted_text,
max_length=512,
padding="max_length",
truncation=True,
return_tensors="pt",
)
device = next(self._defense_model.parameters()).device
input_ids = encoding["input_ids"].to(device)
attention_mask = encoding["attention_mask"].to(device)
return self._defense_model.predict(input_ids, attention_mask)[0]
def analyze_gestalt(
self,
dialogue: list[dict[str, str]],
target_utterance: str,
plutchik_scores: dict[str, float],
ocean_scores: dict[str, float],
max_turns: int = 40,
) -> GestaltState:
"""Fuse signals into a GestaltState."""
plutchik = _validate_scores(
plutchik_scores, PLUTCHIK_EMOTIONS, "plutchik_scores"
)
ocean = _validate_scores(ocean_scores, OCEAN_TRAITS, "ocean_scores")
full_plutchik = {e: plutchik.get(e, 0.0) for e in PLUTCHIK_EMOTIONS}
full_ocean = {t: ocean.get(t, 0.5) for t in OCEAN_TRAITS}
(def_l, def_n, def_c, def_m, def_p) = self._classify_defense(
dialogue, target_utterance, max_turns
)
dom_e, dom_i = _dominant_emotion(full_plutchik)
neuro = full_ocean.get("neuroticism", 0.5)
crisis = _compute_crisis_level(def_l, def_m, dom_e, dom_i, neuro)
behavioral = _behavioral_prediction(def_n, dom_e, crisis, def_m)
directive = _persona_directive(def_l, def_n, def_m)
breakthrough = _breakthrough_score(def_m, self._previous_maturity)
if def_m is not None:
self._previous_maturity = def_m
return GestaltState(
defense_label=def_l,
defense_label_name=def_n,
defense_confidence=def_c,
defense_maturity=def_m,
defense_probabilities=def_p,
plutchik_scores=full_plutchik,
dominant_emotion=dom_e,
dominant_emotion_intensity=dom_i,
ocean_scores=full_ocean,
crisis_level=crisis,
behavioral_prediction=behavioral,
persona_directive=directive,
breakthrough_score=breakthrough,
)
|