Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
File size: 11,812 Bytes
7c3f0ce | 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 | """
QuantumOptimizer — Self-Tuning Engine for the Codette RC+xi Framework.
Inspired by VIVARA Genesis-Omega v2.0, rebuilt as a proper self-tuning system.
The optimizer tracks response quality signals (user engagement, coherence
scores, tension productivity) and adjusts:
- Router confidence thresholds
- Spiderweb parameters (contraction ratio, tension threshold)
- Adapter selection weights
- Multi-perspective synthesis quality
Uses simulated annealing with momentum: explores the parameter space
stochastically but remembers which configurations worked best.
All changes are bounded and reversible. The optimizer logs every
adjustment for full transparency.
"""
from __future__ import annotations
import math
import random
import time
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Tuple
@dataclass
class QualitySignal:
"""A quality signal from a Codette response."""
timestamp: float
adapter: str
coherence: float # Phase coherence at response time
tension: float # Epistemic tension at response time
productivity: float # Tension productivity score
response_length: int # Token count
multi_perspective: bool # Was this a multi-perspective response?
user_continued: bool = True # Did the user continue the conversation?
@dataclass
class TuningState:
"""Current tuning parameters."""
# Router
confidence_threshold: float = 0.4 # Below this, fall back to default
multi_perspective_threshold: float = 0.6 # Above this, force multi-perspective
# Spiderweb
contraction_ratio: float = 0.85
tension_threshold: float = 0.15
entanglement_alpha: float = 0.9
# Adapter weights (0-1 bonus applied to router scores)
adapter_boosts: Dict[str, float] = field(default_factory=dict)
def to_dict(self) -> Dict:
return {
"confidence_threshold": self.confidence_threshold,
"multi_perspective_threshold": self.multi_perspective_threshold,
"contraction_ratio": self.contraction_ratio,
"tension_threshold": self.tension_threshold,
"entanglement_alpha": self.entanglement_alpha,
"adapter_boosts": dict(self.adapter_boosts),
}
@classmethod
def from_dict(cls, data: Dict) -> "TuningState":
state = cls()
for k, v in data.items():
if k == "adapter_boosts":
state.adapter_boosts = dict(v)
elif hasattr(state, k):
setattr(state, k, v)
return state
@dataclass
class OptimizationStep:
"""Record of a single optimization step."""
timestamp: float
parameter: str
old_value: float
new_value: float
reason: str
quality_score: float
class QuantumOptimizer:
"""Self-tuning engine with simulated annealing."""
def __init__(
self,
learning_rate: float = 0.02,
temperature: float = 0.5,
cooling_rate: float = 0.995,
min_signals_before_tuning: int = 5,
):
self.learning_rate = learning_rate
self.temperature = temperature
self.cooling_rate = cooling_rate
self.min_signals = min_signals_before_tuning
self.state = TuningState()
self.best_state = TuningState()
self.best_score = 0.0
self.signals: List[QualitySignal] = []
self.history: List[OptimizationStep] = []
# Running quality metrics
self._quality_window: List[float] = []
self._window_size = 20
def record_signal(self, signal: QualitySignal):
"""Record a quality signal from a Codette response."""
self.signals.append(signal)
# Compute composite quality score
quality = self._compute_quality(signal)
self._quality_window.append(quality)
if len(self._quality_window) > self._window_size:
self._quality_window.pop(0)
# Maybe tune parameters
if len(self.signals) >= self.min_signals:
self._maybe_tune()
def _compute_quality(self, signal: QualitySignal) -> float:
"""Composite quality score from a response signal.
Weights:
- coherence: 30% (high is good — responses make sense)
- productivity: 30% (high is good — tension was resolved productively)
- moderate tension: 20% (sweet spot ~0.3-0.5 is best)
- user_continued: 20% (binary — did they keep talking?)
"""
# Tension is best in the 0.3-0.5 range (productive disagreement)
tension_score = 1.0 - 2.0 * abs(signal.tension - 0.4)
tension_score = max(0.0, tension_score)
quality = (
0.30 * signal.coherence +
0.30 * signal.productivity +
0.20 * tension_score +
0.20 * (1.0 if signal.user_continued else 0.0)
)
return min(max(quality, 0.0), 1.0)
def _maybe_tune(self):
"""Run one optimization step if enough data."""
if len(self._quality_window) < 3:
return
current_quality = sum(self._quality_window) / len(self._quality_window)
# Simulated annealing: accept worse states with decreasing probability
if current_quality > self.best_score:
self.best_score = current_quality
self.best_state = TuningState(**{
k: getattr(self.state, k) for k in vars(self.state)
if not k.startswith('_')
})
elif self.temperature > 0.01:
# Accept worse state with probability exp(-delta/T)
delta = self.best_score - current_quality
accept_prob = math.exp(-delta / max(self.temperature, 0.001))
if random.random() > accept_prob:
# Revert to best known state
self._revert_to_best()
return
# Cool down
self.temperature *= self.cooling_rate
# Pick a parameter to tune based on recent signals
self._tune_one_parameter(current_quality)
def _tune_one_parameter(self, current_quality: float):
"""Tune one parameter based on recent quality signals."""
recent = self.signals[-10:]
# Analyze what needs tuning
avg_coherence = sum(s.coherence for s in recent) / len(recent)
avg_tension = sum(s.tension for s in recent) / len(recent)
avg_productivity = sum(s.productivity for s in recent) / len(recent)
multi_ratio = sum(1 for s in recent if s.multi_perspective) / len(recent)
# Decision: which parameter to adjust
param = None
old_val = 0.0
new_val = 0.0
reason = ""
if avg_coherence < 0.5:
# Low coherence -> increase contraction ratio (tighter belief propagation)
param = "contraction_ratio"
old_val = self.state.contraction_ratio
delta = self.learning_rate * (0.7 - avg_coherence)
new_val = min(0.98, max(0.5, old_val + delta))
reason = f"Low coherence ({avg_coherence:.2f}), tightening propagation"
elif avg_tension < 0.2 and avg_productivity < 0.3:
# Too little tension AND low productivity -> lower confidence threshold
# to allow more multi-perspective responses
param = "multi_perspective_threshold"
old_val = self.state.multi_perspective_threshold
new_val = max(0.3, old_val - self.learning_rate)
reason = f"Low tension+productivity ({avg_tension:.2f}/{avg_productivity:.2f}), encouraging multi-perspective"
elif avg_tension > 0.7:
# Too much tension -> increase tension threshold for convergence
param = "tension_threshold"
old_val = self.state.tension_threshold
new_val = min(0.5, old_val + self.learning_rate * 0.5)
reason = f"High tension ({avg_tension:.2f}), raising convergence threshold"
elif multi_ratio > 0.8 and avg_productivity < 0.4:
# Too many multi-perspective responses but low productivity
param = "multi_perspective_threshold"
old_val = self.state.multi_perspective_threshold
new_val = min(0.8, old_val + self.learning_rate)
reason = f"Multi-perspective overuse ({multi_ratio:.0%}) with low productivity"
# Tune adapter boosts based on which adapters produce best quality
elif len(recent) >= 5:
adapter_quality = {}
for s in recent:
q = self._compute_quality(s)
if s.adapter not in adapter_quality:
adapter_quality[s.adapter] = []
adapter_quality[s.adapter].append(q)
# Boost the best-performing adapter slightly
if adapter_quality:
best_adapter = max(
adapter_quality,
key=lambda a: sum(adapter_quality[a]) / len(adapter_quality[a])
)
param = f"adapter_boost_{best_adapter}"
old_val = self.state.adapter_boosts.get(best_adapter, 0.0)
new_val = min(0.3, old_val + self.learning_rate * 0.5)
self.state.adapter_boosts[best_adapter] = new_val
reason = f"Boosting high-quality adapter: {best_adapter}"
if param and param not in ("adapter_boost_" + a for a in self.state.adapter_boosts):
if hasattr(self.state, param):
setattr(self.state, param, new_val)
if param:
self.history.append(OptimizationStep(
timestamp=time.time(),
parameter=param,
old_value=old_val,
new_value=new_val,
reason=reason,
quality_score=current_quality,
))
def _revert_to_best(self):
"""Revert to the best known tuning state."""
self.state = TuningState(**{
k: getattr(self.best_state, k) for k in vars(self.best_state)
if not k.startswith('_')
})
def get_adapter_boost(self, adapter_name: str) -> float:
"""Get the current boost for an adapter (0.0 = no boost)."""
return self.state.adapter_boosts.get(adapter_name, 0.0)
def get_tuning_report(self) -> Dict:
"""Get current tuning state and recent history."""
recent_quality = (
sum(self._quality_window) / len(self._quality_window)
if self._quality_window else 0.0
)
return {
"current_state": self.state.to_dict(),
"best_score": round(self.best_score, 4),
"current_quality": round(recent_quality, 4),
"temperature": round(self.temperature, 4),
"total_signals": len(self.signals),
"recent_adjustments": [
{
"param": h.parameter,
"old": round(h.old_value, 4),
"new": round(h.new_value, 4),
"reason": h.reason,
}
for h in self.history[-5:]
],
}
def to_dict(self) -> Dict:
"""Serialize for persistence."""
return {
"state": self.state.to_dict(),
"best_score": self.best_score,
"temperature": self.temperature,
"quality_window": self._quality_window,
}
@classmethod
def from_dict(cls, data: Dict) -> "QuantumOptimizer":
opt = cls()
if "state" in data:
opt.state = TuningState.from_dict(data["state"])
opt.best_state = TuningState.from_dict(data["state"])
opt.best_score = data.get("best_score", 0.0)
opt.temperature = data.get("temperature", 0.5)
opt._quality_window = data.get("quality_window", [])
return opt
|