File size: 15,937 Bytes
152f558 | 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 | """Synthetic grounding/faithfulness dataset.
Each example is a short source document plus a one-sentence claim about it, labelled:
0 = grounded (the source supports the claim)
1 = not grounded (unsupported or contradicted by the source)
Design goal: make it HARD, so a trivial word-overlap classifier can't win.
- Grounded claims are often paraphrased or aggregated, so they don't copy source wording.
- Not-grounded claims reuse real entities/numbers from the source, so they look plausible.
Everything is generated from structured facts; no scraped or private text.
"""
from __future__ import annotations
import hashlib
import random
# ---------------------------------------------------------------- fact templates
FIRST = ["Maria", "David", "Aisha", "Chen", "Priya", "Omar", "Sofia", "Liam", "Ingrid", "Raj",
"Elena", "Kofi", "Yuki", "Noah", "Fatima", "Diego", "Hannah", "Wei", "Leila", "Tomas"]
LAST = ["Alvarez", "Okafor", "Nguyen", "Patel", "Andersson", "Haddad", "Rossi", "Kim", "Mbeki",
"Costa", "Larsson", "Reyes", "Tanaka", "Silva", "Novak", "Ibrahim", "Weber", "Owusu"]
CITIES = ["Austin", "Lyon", "Nairobi", "Jakarta", "Porto", "Tallinn", "Medellín", "Pune",
"Leeds", "Osaka", "Calgary", "Valencia", "Accra", "Gdańsk"]
COMPANIES = ["Northwind", "Helios Labs", "Bluefin", "Vireo", "Camberwell Foods", "Opsley",
"Tanager", "Fernweh", "Kestrel Systems", "Marlowe & Co", "Dovetail", "Arcadia Freight"]
PRODUCTS = ["the Atlas tablet", "the Pico earbuds", "the Verge smartwatch", "the Lumen speaker",
"the Drift e-bike", "the Nimbus thermostat", "the Harbor router", "the Sage monitor"]
FIELDS = ["marine biology", "supply-chain logistics", "renewable energy", "medieval history",
"behavioral economics", "materials science", "urban planning", "epidemiology"]
def _person(rng):
return f"{rng.choice(FIRST)} {rng.choice(LAST)}"
def make_facts(rng) -> dict:
kind = rng.choice(["company", "product", "person", "study"])
if kind == "company":
f = {
"kind": "company",
"name": rng.choice(COMPANIES),
"city": rng.choice(CITIES),
"founded": rng.randint(1985, 2021),
"employees": rng.choice([42, 85, 120, 240, 510, 900, 1300, 2600]),
"revenue": rng.choice([3, 7, 12, 24, 55, 88, 140]), # $M
"sector": rng.choice(["logistics", "consumer hardware", "food distribution", "fintech", "clean energy"]),
"ceo": _person(rng),
"growth": rng.choice([6, 11, 18, 27, 34]), # % YoY
}
elif kind == "product":
f = {
"kind": "product",
"name": rng.choice(PRODUCTS),
"maker": rng.choice(COMPANIES),
"price": rng.choice([49, 89, 129, 199, 299, 449]),
"battery": rng.choice([6, 9, 12, 18, 24, 36]), # hours
"weight": rng.choice([120, 180, 240, 320, 460]), # grams
"released": rng.choice(["March", "June", "September", "November"]),
"year": rng.randint(2022, 2026),
"color_count": rng.choice([2, 3, 4, 5]),
}
elif kind == "person":
f = {
"kind": "person",
"name": _person(rng),
"city": rng.choice(CITIES),
"field": rng.choice(FIELDS),
"years": rng.choice([4, 7, 11, 15, 20]),
"degree": rng.choice(["a master's", "a PhD", "a bachelor's"]),
"awards": rng.choice([0, 1, 2, 3]),
"team": rng.choice([3, 6, 9, 14]),
}
else:
f = {
"kind": "study",
"field": rng.choice(FIELDS),
"n": rng.choice([48, 120, 350, 900, 2100]),
"duration": rng.choice([6, 12, 18, 24, 36]), # months
"effect": rng.choice([8, 14, 22, 31, 45]), # % change
"city": rng.choice(CITIES),
"lead": _person(rng),
"year": rng.randint(2019, 2026),
}
return f
# ---------------------------------------------------------------- source rendering
def render_source(f, rng) -> str:
k = f["kind"]
if k == "company":
return (f"{f['name']} is a {f['sector']} company based in {f['city']}, founded in {f['founded']}. "
f"It employs around {f['employees']} people and reported annual revenue of about "
f"${f['revenue']} million last year. {f['ceo']} serves as chief executive. "
f"Revenue grew {f['growth']}% year over year.")
if k == "product":
return (f"{f['name'].capitalize()} is made by {f['maker']} and sells for ${f['price']}. "
f"It offers up to {f['battery']} hours of battery life and weighs {f['weight']} grams. "
f"It was released in {f['released']} {f['year']} and comes in {f['color_count']} colors.")
if k == "person":
return (f"{f['name']} is a researcher in {f['field']} based in {f['city']}, with {f['years']} years "
f"of experience. They hold {f['degree']} degree and lead a team of {f['team']}. "
f"They have received {f['awards']} industry awards.")
return (f"A study in {f['field']} led by {f['lead']} followed {f['n']} participants over {f['duration']} months "
f"in {f['city']}, completed in {f['year']}. It reported a {f['effect']}% change in the measured outcome.")
# ---------------------------------------------------------------- grounded claims (often paraphrased)
NUM_WORDS = {2: "two", 3: "three", 4: "four", 5: "five", 6: "six"}
def grounded_claim(f, rng) -> str:
k = f["kind"]
opts = []
if k == "company":
opts += [
f"The firm operates in the {f['sector']} space.",
f"{f['name']} has its headquarters in {f['city']}.",
f"The company has been around since {f['founded']}.",
f"Its workforce numbers roughly {f['employees']}.",
f"{f['ceo']} is the person at the top of {f['name']}.",
f"Sales climbed by {f['growth']}% compared with the prior year.",
f"Last year's turnover came in near ${f['revenue']} million.",
f"The business has grown over {2026 - f['founded']} years of operation.",
]
elif k == "product":
opts += [
f"{f['maker']} is the manufacturer behind this device.",
f"It carries a price tag of ${f['price']}.",
f"On a full charge it can run for as long as {f['battery']} hours.",
f"The unit tips the scales at {f['weight']} grams.",
f"It first went on sale in {f['released']} of {f['year']}.",
f"Buyers can choose from {NUM_WORDS.get(f['color_count'], f['color_count'])} color options.",
f"It is a fairly light device at well under a kilogram.",
]
elif k == "person":
opts += [
f"Their work centers on {f['field']}.",
f"They are based in {f['city']}.",
f"They bring {f['years']} years of experience to the role.",
f"They oversee a group of {f['team']} people.",
f"They completed {f['degree']} degree.",
(f"They have not won any industry awards." if f["awards"] == 0
else f"They have picked up {f['awards']} awards in the industry."),
]
else:
opts += [
f"The research focused on {f['field']}.",
f"{f['lead']} headed the project.",
f"It enrolled {f['n']} participants.",
f"The work ran for {f['duration']} months.",
f"Researchers observed a {f['effect']}% shift in the outcome.",
f"The study took place in {f['city']}.",
f"It wrapped up in {f['year']}.",
]
return rng.choice(opts)
# ---------------------------------------------------------------- not-grounded claims (plausible, reuse real entities)
def _other(rng, pool, not_this):
pool = [x for x in pool if x != not_this]
return rng.choice(pool)
def not_grounded_claim(f, rng) -> tuple[str, str]:
"""Return (claim, subtype). Subtypes: invented, altered_number, swapped_entity, contradicted, overgeneral."""
k = f["kind"]
builders = []
if k == "company":
builders += [
(f"{f['name']} recently opened a second office in {_other(rng, CITIES, f['city'])}.", "invented"),
(f"The company employs about {rng.choice([x for x in [42,85,120,240,510,900,1300,2600] if x != f['employees']])} people.", "altered_number"),
(f"{f['name']} is led by {_person(rng)}.", "swapped_entity"),
(f"Revenue actually fell {f['growth']}% last year.", "contradicted"),
(f"The firm is the largest {f['sector']} company in the country.", "overgeneral"),
(f"It was founded in {rng.choice([x for x in range(1985,2022) if abs(x-f['founded'])>3])}.", "altered_number"),
]
elif k == "product":
builders += [
(f"{f['name'].capitalize()} is waterproof to 50 meters.", "invented"),
(f"It costs ${_other(rng, [49,89,129,199,299,449], f['price'])}.", "altered_number"),
(f"It is manufactured by {_other(rng, COMPANIES, f['maker'])}.", "swapped_entity"),
(f"Battery life tops out at just {max(1, f['battery']-5)} hours.", "contradicted"),
(f"It is the best-selling device in its category.", "overgeneral"),
(f"It launched in {_other(rng, ['March','June','September','November'], f['released'])} {f['year']}.", "altered_number"),
]
elif k == "person":
builders += [
(f"They previously founded a startup in {_other(rng, CITIES, f['city'])}.", "invented"),
(f"They manage a team of {_other(rng, [3,6,9,14], f['team'])}.", "altered_number"),
(f"Their research is primarily in {_other(rng, FIELDS, f['field'])}.", "swapped_entity"),
((f"They have won {rng.choice([1,2,3])} awards." if f['awards']==0
else "They have never received any industry award."), "contradicted"),
(f"They are among the most cited researchers in their field.", "overgeneral"),
]
else:
builders += [
(f"The study was funded by a government grant.", "invented"),
(f"It followed {_other(rng, [48,120,350,900,2100], f['n'])} participants.", "altered_number"),
(f"It was led by {_person(rng)}.", "swapped_entity"),
(f"The intervention showed no measurable effect.", "contradicted"),
(f"It is the largest study of its kind to date.", "overgeneral"),
(f"The research was conducted in {_other(rng, CITIES, f['city'])}.", "swapped_entity"),
]
return rng.choice(builders)
def grounded_hard(f, rng):
"""Supported, but via inference/aggregation/paraphrase — low lexical overlap, looks 'novel'."""
k = f["kind"]
opts = []
if k == "company":
decades = (2026 - f["founded"]) // 10
opts += [
(f"The business has been operating for more than {max(1,decades)} decade" + ("s" if decades!=1 else "") + ".", "inference"),
("It is a mid-sized organization rather than a tiny startup." if f["employees"] >= 240
else "It remains a relatively small outfit.", "inference"),
("Revenue moved in a positive direction over the past year." if f["growth"] > 0 else "Revenue slipped over the year.", "paraphrase"),
(f"Hundreds of staff are on its payroll." if f["employees"] >= 100 else "It has only a few dozen staff.", "inference"),
]
elif k == "product":
opts += [
("It sits at the budget end of the range." if f["price"] <= 129 else "It is positioned as a premium option.", "inference"),
("A single charge comfortably lasts a full workday." if f["battery"] >= 9 else "Its battery barely covers a few hours.", "inference"),
("It is light enough to carry around easily." if f["weight"] <= 320 else "It is on the heavier side.", "inference"),
("It arrived on the market in the latter half of the year." if f["released"] in ("September","November") else "It launched in the first half of the year.", "inference"),
]
elif k == "person":
opts += [
("They are a seasoned expert with over a decade in the field." if f["years"] >= 11 else "They are relatively early in their career.", "inference"),
("They have advanced academic training." if f["degree"] in ("a PhD","a master's") else "Their background is at the undergraduate level.", "paraphrase"),
("They manage a small group rather than a large department." if f["team"] <= 9 else "They run a sizeable team.", "inference"),
]
else:
opts += [
("The trial ran for well over a year." if f["duration"] > 12 else "The trial was relatively short.", "inference"),
("It was a large-scale study by sample size." if f["n"] >= 900 else "It was a modest-sized study.", "inference"),
("The effect observed was substantial." if f["effect"] >= 22 else "The effect observed was fairly small.", "inference"),
]
c, sub = rng.choice(opts)
return c, sub
def not_grounded_hard(f, rng):
"""Unsupported, but reuses real numbers/entities so lexical support stays high."""
k = f["kind"]
opts = []
if k == "company":
opts += [
(f"Its ${f['revenue']} million in revenue makes it the market leader in {f['sector']}.", "unsupported_relation"),
(f"With {f['employees']} employees, it is the largest employer in {f['city']}.", "unsupported_relation"),
(f"{f['ceo']} founded the company back in {f['founded']}.", "unsupported_relation"),
(f"Its {f['growth']}% growth is the fastest in the {f['sector']} sector.", "unsupported_relation"),
]
elif k == "product":
opts += [
(f"At ${f['price']}, {f['name']} is the cheapest option on the market.", "unsupported_relation"),
(f"Its {f['battery']}-hour battery is the longest in its class.", "unsupported_relation"),
(f"{f['maker']} designed {f['name']} specifically for professional athletes.", "unsupported_relation"),
(f"The {f['color_count']} color options were chosen after extensive customer surveys.", "unsupported_relation"),
]
elif k == "person":
opts += [
(f"Their {f['years']} years in {f['field']} make them the top authority in {f['city']}.", "unsupported_relation"),
(f"They lead their team of {f['team']} remotely from abroad.", "unsupported_relation"),
(f"Their {f['degree']} degree was earned at an Ivy League university.", "unsupported_relation"),
]
else:
opts += [
(f"The {f['effect']}% effect proves the intervention is cost-effective.", "unsupported_relation"),
(f"{f['lead']} has run similar studies for over twenty years.", "unsupported_relation"),
(f"The {f['n']} participants were all recruited from {f['city']} hospitals.", "unsupported_relation"),
]
return rng.choice(opts)
def make_example(idx: int, seed: int = 20261113) -> dict:
rng = random.Random(seed * 7_000_003 + idx)
f = make_facts(rng)
source = render_source(f, rng)
roll = rng.random()
if roll < 0.32:
claim, subtype, label = grounded_claim(f, rng), "grounded", 0
elif roll < 0.50:
claim, subtype = grounded_hard(f, rng); label = 0
elif roll < 0.82:
claim, subtype = not_grounded_claim(f, rng); label = 1
else:
claim, subtype = not_grounded_hard(f, rng); label = 1
uid = hashlib.sha1(f"{seed}-{idx}".encode()).hexdigest()[:12]
return {
"id": f"grnd-{uid}",
"domain": f["kind"],
"source": source,
"claim": claim,
"label": label, # 1 = not grounded
"subtype": subtype,
}
|