Spaces:
Sleeping
Sleeping
File size: 7,939 Bytes
09f7d63 | 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 | from __future__ import annotations
import math
from collections import Counter
from typing import Dict, List, Optional
from viral_script_engine.memory.creator_history import CreatorHistoryBuffer, EpisodeMemory
_REWARD_KEYS = [
"r1_hook_strength",
"r2_coherence",
"r3_cultural_alignment",
"r4_debate_resolution",
"r5_defender_preservation",
"r6_safety",
"r7_originality",
"r8_persona_fit",
"r9_platform_pacing",
]
_DELTA_THRESHOLD = 0.05
class MemoryCompressor:
"""
Compresses a completed episode into a structured EpisodeMemory.
Called at the end of every episode, before the next reset().
Zero LLM calls — all compression is rule-based.
"""
def compress(self, episode_log: dict, episode_number: int) -> EpisodeMemory:
"""
episode_log fields expected:
episode_id, niche, platform, first_critique_claims,
actions_taken, initial_reward_components, final_reward_components,
final_total_reward
"""
episode_id = episode_log.get("episode_id", "unknown")
niche = episode_log.get("niche", "unknown")
platform = episode_log.get("platform", "unknown")
actions_taken: List[str] = episode_log.get("actions_taken", [])
initial_rc: dict = episode_log.get("initial_reward_components", {})
final_rc: dict = episode_log.get("final_reward_components", {})
final_total = episode_log.get("final_total_reward", 0.0)
# 1. dominant_flaw: most common critique_class from first-step claims
first_claims = episode_log.get("first_critique_claims", [])
if first_claims:
counts = Counter(
c.get("critique_class", "unknown") for c in first_claims
)
dominant_flaw = counts.most_common(1)[0][0]
else:
dominant_flaw = "hook_weakness"
# 2. what_worked / what_didnt — reward components with significant delta
what_worked: List[str] = []
what_didnt: List[str] = []
for key in _REWARD_KEYS:
init_val = initial_rc.get(key)
final_val = final_rc.get(key)
if init_val is None or final_val is None:
continue
delta = final_val - init_val
if delta > _DELTA_THRESHOLD:
what_worked.append(key)
elif delta < -_DELTA_THRESHOLD:
what_didnt.append(key)
# 3. key_learning — rule-based template
most_used_action = (
Counter(actions_taken).most_common(1)[0][0] if actions_taken else "no_action"
)
worked_str = what_worked[0] if what_worked else "no component"
didnt_str = what_didnt[0] if what_didnt else "no regressions"
key_learning = (
f"Fixed {dominant_flaw} using {most_used_action}. "
f"{worked_str} improved, {didnt_str}."
)
return EpisodeMemory(
episode_id=episode_id,
episode_number=episode_number,
script_niche=niche,
platform=platform,
dominant_flaw=dominant_flaw,
actions_taken=actions_taken,
what_worked=what_worked,
what_didnt=what_didnt,
final_total_reward=final_total,
key_learning=key_learning,
)
def update_buffer(
self,
existing_buffer: Optional[CreatorHistoryBuffer],
new_memory: EpisodeMemory,
creator_id: str,
) -> CreatorHistoryBuffer:
"""
Adds new_memory to the buffer, maintaining a sliding window of 5.
Recomputes all aggregate stats.
"""
if existing_buffer is None:
episodes: List[EpisodeMemory] = []
total = 0
else:
episodes = list(existing_buffer.recent_episodes)
total = existing_buffer.total_episodes
episodes.append(new_memory)
if len(episodes) > 5:
episodes = episodes[-5:] # keep last 5
total += 1
# recurring_weak_points: dominant_flaw in >= 3 of last 5
flaw_counts = Counter(ep.dominant_flaw for ep in episodes)
recurring_weak_points = [
flaw for flaw, cnt in flaw_counts.items() if cnt >= 3
]
# recurring_strong_points: reward component >= 0.7 in >= 4 of last 5
recurring_strong_points = self._compute_strong_points(episodes)
# most_effective_action: action_type with highest avg final_total_reward
most_effective_action = self._compute_most_effective_action(episodes)
# voice_stability_score: 1 - std_dev of r3 across episodes (inverted, clamped)
voice_stability_score = self._compute_voice_stability(episodes)
# improvement_trend: slope of final_total_reward
improvement_trend = self._compute_trend(episodes)
return CreatorHistoryBuffer(
creator_id=creator_id,
total_episodes=total,
recent_episodes=episodes,
recurring_weak_points=recurring_weak_points,
recurring_strong_points=recurring_strong_points,
most_effective_action=most_effective_action,
voice_stability_score=voice_stability_score,
improvement_trend=improvement_trend,
)
# ------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------
def _compute_strong_points(self, episodes: List[EpisodeMemory]) -> List[str]:
"""Reward components consistently >= 0.7 in >= 4 of last 5 episodes."""
if not episodes:
return []
# We only know what_worked from EpisodeMemory — approximate by checking
# which components appear in what_worked across >= 4 episodes
counts: Dict[str, int] = {}
for ep in episodes:
for comp in ep.what_worked:
counts[comp] = counts.get(comp, 0) + 1
threshold = max(4, len(episodes) - 1) if len(episodes) >= 4 else len(episodes)
return [comp for comp, cnt in counts.items() if cnt >= threshold]
def _compute_most_effective_action(self, episodes: List[EpisodeMemory]) -> Optional[str]:
"""Action type with highest average final_total_reward across episodes it appeared in."""
if not episodes:
return None
action_rewards: Dict[str, List[float]] = {}
for ep in episodes:
for action in set(ep.actions_taken):
action_rewards.setdefault(action, []).append(ep.final_total_reward)
if not action_rewards:
return None
return max(action_rewards, key=lambda a: sum(action_rewards[a]) / len(action_rewards[a]))
def _compute_voice_stability(self, episodes: List[EpisodeMemory]) -> float:
"""Stability of R3 inferred from whether r3_cultural_alignment was in what_didnt.
A proxy: episodes where R3 did NOT regress count toward stability."""
if not episodes:
return 1.0
stable_count = sum(
1 for ep in episodes if "r3_cultural_alignment" not in ep.what_didnt
)
return stable_count / len(episodes)
def _compute_trend(self, episodes: List[EpisodeMemory]) -> str:
"""Slope of final_total_reward across the episode window."""
if len(episodes) < 2:
return "plateauing"
rewards = [ep.final_total_reward for ep in episodes]
n = len(rewards)
x_mean = (n - 1) / 2.0
y_mean = sum(rewards) / n
numerator = sum((i - x_mean) * (rewards[i] - y_mean) for i in range(n))
denominator = sum((i - x_mean) ** 2 for i in range(n))
if denominator == 0:
return "plateauing"
slope = numerator / denominator
if slope > 0.02:
return "improving"
elif slope < -0.02:
return "declining"
return "plateauing"
|