Instructions to use Falln87/clerk-memory with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Falln87/clerk-memory with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
CLERK synthetic evolving-session generator: personas, updates/supersessions/tombstones, gold ledger transitions, seeded
Browse files- clerk/generator.py +178 -0
clerk/generator.py
ADDED
|
@@ -0,0 +1,178 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""CLERK synthetic evolving-session generator.
|
| 2 |
+
|
| 3 |
+
Generates personas with evolving facts across multiple dialogue sessions.
|
| 4 |
+
Because every fact event (ADD / UPDATE / TOMBSTONE) is generated
|
| 5 |
+
programmatically, the gold ledger state after every session is known by
|
| 6 |
+
construction, which yields exact supervision for the consolidation policy β
|
| 7 |
+
no LLM judge anywhere in the pipeline.
|
| 8 |
+
|
| 9 |
+
Everything is seeded: two runs with the same --seed produce byte-identical
|
| 10 |
+
data.
|
| 11 |
+
|
| 12 |
+
Author: Justin Wolcott (fallnai-research.org)
|
| 13 |
+
License: Apache-2.0
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
from __future__ import annotations
|
| 17 |
+
|
| 18 |
+
import argparse
|
| 19 |
+
import json
|
| 20 |
+
import random
|
| 21 |
+
from typing import Any, Dict, List, Optional, Tuple
|
| 22 |
+
|
| 23 |
+
from clerk.common import (
|
| 24 |
+
LEDGER_BUDGET_DEFAULT,
|
| 25 |
+
Slot,
|
| 26 |
+
apply_ops,
|
| 27 |
+
empty_ledger,
|
| 28 |
+
enforce_budget,
|
| 29 |
+
serialize_ledger,
|
| 30 |
+
salience,
|
| 31 |
+
)
|
| 32 |
+
|
| 33 |
+
# ---------------------------------------------------------------- predicates
|
| 34 |
+
|
| 35 |
+
# (predicate key, human phrase, value pool, statement template, update
|
| 36 |
+
# template, tombstone template)
|
| 37 |
+
PREDICATES: List[Dict[str, Any]] = [
|
| 38 |
+
{"p": "job", "h": "job", "v": ["barista", "chef", "nurse", "high school teacher", "librarian", "electrician", "accountant", "software engineer", "pharmacist", "bus driver"],
|
| 39 |
+
"add": "Big news β I started a new job as a {v}.", "upd": "I actually switched jobs again. I'm a {v} now.", "tmb": "I left my job β between things right now."},
|
| 40 |
+
{"p": "city", "h": "city", "v": ["Denver", "Portland", "Nashville", "Austin", "Madison", "Boise", "Tucson", "Salem", "Reno", "Spokane"],
|
| 41 |
+
"add": "I just moved to {v}!", "upd": "We relocated again β I'm in {v} now.", "tmb": "I'm between cities at the moment, staying with family."},
|
| 42 |
+
{"p": "pet_kind", "h": "pet", "v": ["cat", "dog", "rabbit", "parakeet", "hamster", "iguana"],
|
| 43 |
+
"add": "I adopted a {v}.", "upd": "Long story, but I rehomed the {v_old} and got a {v} instead.", "tmb": "My pet passed away last month. It's been rough."},
|
| 44 |
+
{"p": "partner", "h": "partner", "v": ["Sam", "Jordan", "Priya", "Diego", "Nina", "Marcus", "Ellie", "Omar"],
|
| 45 |
+
"add": "I've been seeing someone β their name is {v}.", "upd": "That ended, and I'm dating {v} now.", "tmb": "We broke up, so I'm single again."},
|
| 46 |
+
{"p": "car", "h": "car", "v": ["a red Civic", "a blue Outback", "a gray Silverado", "a used Miata", "an old Corolla", "a white Jetta"],
|
| 47 |
+
"add": "I finally bought a car β {v}.", "upd": "I traded it in for {v}.", "tmb": "I sold the car β biking everywhere now."},
|
| 48 |
+
{"p": "hobby", "h": "main hobby", "v": ["bouldering", "watercolor painting", "chess", "birdwatching", "pottery", "kickboxing", "sourdough baking", "trail running"],
|
| 49 |
+
"add": "I've gotten really into {v}.", "upd": "I switched hobbies β now it's all about {v}.", "tmb": "I had to drop the hobby, no time lately."},
|
| 50 |
+
{"p": "favorite_food", "h": "favorite food", "v": ["pad thai", "dumplings", "tacos", "ramen", "lasagna", "pho", "shawarma", "pancakes"],
|
| 51 |
+
"add": "If I had to pick a favorite food, it's {v}.", "upd": "I've changed my favorite food to {v}, believe it or not.", "tmb": "I can't eat that anymore β no favorite food these days."},
|
| 52 |
+
{"p": "gym", "h": "gym", "v": ["RiseFit", "IronWorks", "FlexCity", "PeakBarre", "the Strength Lab", "Northside Athletic"],
|
| 53 |
+
"add": "I joined a new gym called {v}.", "upd": "I switched gyms β now I'm at {v}.", "tmb": "I cancelled my gym membership."},
|
| 54 |
+
{"p": "boss", "h": "boss", "v": ["Dana", "Victor", "Renee", "Kofi", "Beatrice", "Hank"],
|
| 55 |
+
"add": "My new boss's name is {v}.", "upd": "There's a new boss over me now: {v}.", "tmb": "My boss left and they haven't replaced them."},
|
| 56 |
+
{"p": "coworker", "h": "closest coworker", "v": ["Lena", "Theo", "Yusuf", "Marge", "Pete", "Iris"],
|
| 57 |
+
"add": "I've become friends with a coworker named {v}.", "upd": "On a new team now β my closest coworker there is {v}.", "tmb": "That coworker transferred away."},
|
| 58 |
+
{"p": "project", "h": "big work project", "v": ["the website redesign", "the database migration", "the product launch", "the security audit", "the onboarding revamp"],
|
| 59 |
+
"add": "I'm leading {v} at work.", "upd": "I got moved off that and onto {v}.", "tmb": "The project got cancelled, honestly."},
|
| 60 |
+
{"p": "deadline", "h": "big deadline", "v": ["March 14", "April 2", "May 9", "June 21", "July 3", "August 30"],
|
| 61 |
+
"add": "My big deadline is {v}.", "upd": "The deadline moved β it's {v} now.", "tmb": "The deadline got pushed indefinitely."},
|
| 62 |
+
{"p": "instrument", "h": "instrument", "v": ["guitar", "violin", "accordion", "cello", "drums", "trumpet"],
|
| 63 |
+
"add": "I've been learning the {v}.", "upd": "I switched instruments β playing the {v} now.", "tmb": "I gave up the instrument, wrists couldn't take it."},
|
| 64 |
+
{"p": "team", "h": "sports team I follow", "v": ["the Larks", "the Comets", "the Harbors", "the Northside Pistons", "the Rovers"],
|
| 65 |
+
"add": "I've started following {v}.", "upd": "I switched allegiances β it's {v} now.", "tmb": "I stopped following sports entirely."},
|
| 66 |
+
{"p": "allergy", "h": "allergy", "v": ["shellfish", "peanuts", "latex", "penicillin", "kiwi"],
|
| 67 |
+
"add": "Heads up: I'm allergic to {v}.", "upd": "New allergy diagnosis β {v} this time.", "tmb": "Good news, the allergy testing came back clear."},
|
| 68 |
+
{"p": "vacation_plan", "h": "next vacation spot", "v": ["Lisbon", "Kyoto", "Banff", "Oaxaca", "the Azores", "Marrakesh"],
|
| 69 |
+
"add": "I'm planning a trip to {v}.", "upd": "Change of plans β we're going to {v} instead.", "tmb": "The trip is off."},
|
| 70 |
+
{"p": "hobby_class", "h": "evening class", "v": ["a pottery class", "a Spanish class", "a jazz piano class", "a photography class", "a welding class"],
|
| 71 |
+
"add": "I signed up for {v}.", "upd": "I dropped that and enrolled in {v}.", "tmb": "I withdrew from the class."},
|
| 72 |
+
{"p": "neighbor", "h": "neighbor", "v": ["Mr. Alvarez", "Ms. Chen", "the Kowalskis", "Pastor Jim", "Dr. Patel"],
|
| 73 |
+
"add": "My new neighbor is {v}.", "upd": "They moved out β my neighbor now is {v}.", "tmb": "My neighbor moved away recently."},
|
| 74 |
+
]
|
| 75 |
+
|
| 76 |
+
TRAIN_NAMES = ["Alex", "Bailey", "Casey", "Drew", "Emery", "Frankie", "Gray", "Harper",
|
| 77 |
+
"Indigo", "Jesse", "Kai", "Logan", "Marley", "Noel", "Oakley", "Parker",
|
| 78 |
+
"Quinn", "Reese", "Sage", "Tatum", "Undra", "Val", "Wren", "Xander"]
|
| 79 |
+
TEST_NAMES = ["Ari", "Bellamy", "Crew", "Dallas", "East", "Flynn", "Greer", "Hollis",
|
| 80 |
+
"Ira", "Jules", "Kit", "Lennon", "Morgan", "Nico", "Onyx", "Perry",
|
| 81 |
+
"Quince", "Rory", "Sailor", "Tao", "Uri", "Vesper", "Wilder", "Yael"]
|
| 82 |
+
|
| 83 |
+
CHATTER = [
|
| 84 |
+
"Anyway β how's your week going?",
|
| 85 |
+
"Sorry, I just needed to vent a little.",
|
| 86 |
+
"This weather is something else lately.",
|
| 87 |
+
"Did you catch the game last night?",
|
| 88 |
+
"I should probably get more sleep.",
|
| 89 |
+
"Anyway, enough about that.",
|
| 90 |
+
"My commute was ridiculous today.",
|
| 91 |
+
"Coffee count today: three. Send help.",
|
| 92 |
+
]
|
| 93 |
+
ACKS = ["Got it β noted.", "That's good to hear.", "Thanks for the update!",
|
| 94 |
+
"Ah, interesting.", "Makes sense.", "Oh wow, okay.", "Sounds like a plan.",
|
| 95 |
+
"Noted, thanks for telling me."]
|
| 96 |
+
|
| 97 |
+
Q_TEMPLATES = {
|
| 98 |
+
"recent": ["Do you remember what my {h} is?", "Quick quiz: what's my {h} again?"],
|
| 99 |
+
"stale": ["Do you still remember my {h} from before?", "You remember what my {h} is, right?"],
|
| 100 |
+
"superseded": ["Wait β what is my {h} these days?", "Remind me: what's my current {h}?"],
|
| 101 |
+
"negated": ["Do I still have a {h}?", "What about my {h} β still a thing?"],
|
| 102 |
+
"unknown": ["What's my {h}?", "Do you happen to know my {h}?"],
|
| 103 |
+
"temporal": ["What was my {h} back then, before it changed?", "What did my {h} used to be?"],
|
| 104 |
+
}
|
| 105 |
+
|
| 106 |
+
ANSWER_OK = {
|
| 107 |
+
"value": "{v}",
|
| 108 |
+
"negated": "You don't have that anymore β you mentioned it's no longer the case.",
|
| 109 |
+
"unknown": "I don't know β you haven't told me about that yet.",
|
| 110 |
+
"temporal": "{v}",
|
| 111 |
+
}
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def _fill(t: str, v: Optional[str] = None, h: Optional[str] = None, s: Optional[str] = None) -> str:
|
| 115 |
+
out = t
|
| 116 |
+
if v is not None:
|
| 117 |
+
out = out.replace("{v}", v)
|
| 118 |
+
if h is not None:
|
| 119 |
+
out = out.replace("{h}", h)
|
| 120 |
+
if s is not None:
|
| 121 |
+
out = out.replace("{s}", s)
|
| 122 |
+
return out
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
# ---------------------------------------------------------------- generation
|
| 126 |
+
|
| 127 |
+
def make_persona(rng: random.Random, name: str, n_init: int = 8) -> Dict[str, Any]:
|
| 128 |
+
preds = rng.sample(PREDICATES, n_init)
|
| 129 |
+
facts = []
|
| 130 |
+
for spec in preds:
|
| 131 |
+
facts.append({
|
| 132 |
+
"pred": spec["p"], "value": rng.choice(spec["v"]),
|
| 133 |
+
"born_session": 0, "state": "active", # active | updated | tombstoned
|
| 134 |
+
"history": [],
|
| 135 |
+
})
|
| 136 |
+
return {"name": name, "facts": facts}
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def find_fact(persona: Dict[str, Any], pred: str) -> Optional[Dict[str, Any]]:
|
| 140 |
+
for f in persona["facts"]:
|
| 141 |
+
if f["pred"] == pred:
|
| 142 |
+
return f
|
| 143 |
+
return None
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
def spec_of(pred: str) -> Dict[str, Any]:
|
| 147 |
+
return next(p for p in PREDICATES if p["p"] == pred)
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def render_session(
|
| 151 |
+
rng: random.Random,
|
| 152 |
+
persona: Dict[str, Any],
|
| 153 |
+
session_idx: int,
|
| 154 |
+
n_add: int, n_upd: int, n_tmb: int, n_q: int,
|
| 155 |
+
) -> Tuple[List[Dict[str, str]], List[Dict[str, Any]], List[Dict[str, Any]]]:
|
| 156 |
+
"""Render one session. Returns (turns, write_events, qas).
|
| 157 |
+
|
| 158 |
+
write_events: [{"kind": ADD/UPDATE/TOMBSTONE, "pred":..., "value":...}]
|
| 159 |
+
qas: [{"q":..., "a":..., "type":..., "pred":...}]
|
| 160 |
+
"""
|
| 161 |
+
name = persona["name"]
|
| 162 |
+
turns: List[Dict[str, Any]] = []
|
| 163 |
+
events: List[Dict[str, Any]] = []
|
| 164 |
+
|
| 165 |
+
# --- pick events
|
| 166 |
+
updatable = [f for f in persona["facts"] if f["state"] == "active"
|
| 167 |
+
and f["born_session"] < session_idx]
|
| 168 |
+
spec_pool = [p for p in PREDICATES if find_fact(persona, p["p"]) is None]
|
| 169 |
+
|
| 170 |
+
for _ in range(n_add):
|
| 171 |
+
if not spec_pool_left(spec_pool_left_spec := spec_pool_left := None) and False:
|
| 172 |
+
pass # placeholder removed
|
| 173 |
+
del spec_pool_left
|
| 174 |
+
|
| 175 |
+
for _ in range(n_add):
|
| 176 |
+
if not spec_pool_left_check(spec_pool_left_spec := None) and False:
|
| 177 |
+
pass
|
| 178 |
+
return [], [], []
|