tempbench / code /build_benchmark.py
Guen's picture
Add the deterministic construction pipeline
210b340 verified
Raw
History Blame Contribute Delete
41.8 kB
"""
TempBench Builder
==========================
Constructs a 10K-question multi-hop temporal QA benchmark (see the TempBench paper).
Implements the full 6-stage pipeline:
1. Question generation from KG triples and templates
2. Composability filtering (validate temporal consistency)
3. Answer uniqueness filtering (discard ambiguous questions)
4. MinHash deduplication (Jaccard similarity threshold)
5. Subgraph construction (S_star, S_dist, S_stale)
6. Train/dev/test split (stratified by complexity)
Output: JSONL benchmark with one question per line, including supporting subgraphs.
Usage:
python build_benchmark.py --kg_path kg.json --output_dir ./benchmark/ --target_n 10000
"""
from __future__ import annotations
import argparse
import hashlib
import json
import math
import random
import sys
from collections import defaultdict
from dataclasses import dataclass, asdict
from pathlib import Path
from typing import Dict, List, Optional, Set, Tuple
from indexer import (
TemporalKGIndexer,
Triple,
ValidityWindow,
)
# ---------------------------------------------------------------------------
# Question generation templates
# ---------------------------------------------------------------------------
QUESTION_TEMPLATES = {
# The TempBench paper cites the TimelineKGQA taxonomy for four operator types.
# Each operator takes a compositional chain "path" (= anchor + hop
# relations, joined by "'s") and attaches an operator-specific temporal
# qualifier. See benchmark-design-decisions.md §1 for semantics.
"point_in_time": [
"In {year}, what was the {final_relation} of {path}?",
"What was the {final_relation} of {path} in {year}?",
"{path}'s {final_relation} in {year} was?",
],
"before_after": [
"Who was the {final_relation} of {path} {qualifier} {sibling}?",
"{qualifier_cap} {sibling}, who was the {final_relation} of {path}?",
],
"interval": [
"In what year was the {final_relation} of {path} equal to {terminal}?",
"When was {terminal} the {final_relation} of {path}?",
],
"sequence": [
"After {ref_subject} {ref_relation} {ref_object}, what was the {final_relation} of {path}?",
"What was the {final_relation} of {path} after {ref_subject} {ref_relation} {ref_object}?",
],
}
def _render_composition_path(anchor: str, relations: List[str]) -> str:
"""
Render the nested possessive prefix for compositional multi-hop questions.
Example:
anchor = "Obama", relations = ["spouse", "country"]
returns: "Obama's spouse's country"
Used as the {path} placeholder for the last-but-one hop; the final
relation becomes the {final_relation} placeholder.
"""
parts = [anchor] + relations
return "'s ".join(parts)
# ---------------------------------------------------------------------------
# Data structures
# ---------------------------------------------------------------------------
@dataclass
class Candidate:
"""An intermediate candidate question before filtering."""
id: str
question: str
t_query: float
answer: str
complexity: str # "1hop", "2hop", "3plus"
operator_type: str # "point_in_time", "before_after", "interval", "sequence"
question_type: str # "explicit", "implicit", "ordinal"
gold_chain: List[Triple]
@dataclass
class BenchmarkQuestion:
"""A final, fully-validated benchmark question."""
id: str
question: str
t_query: float
answer: str
complexity: str
operator_type: str
question_type: str
S_star: List[Dict] # Gold chain
S_dist: List[Dict] # Distractor chain
S_stale: List[Dict] # Stale-fact chain
split: str # "train", "dev", "test"
def to_jsonl_dict(self) -> Dict:
"""Convert to dict for JSONL serialization."""
return {
"id": self.id,
"question": self.question,
"t_query": self.t_query,
"answer": self.answer,
"complexity": self.complexity,
"operator_type": self.operator_type,
"question_type": self.question_type,
"S_star": self.S_star,
"S_dist": self.S_dist,
"S_stale": self.S_stale,
"split": self.split,
}
# ---------------------------------------------------------------------------
# Stage 1: Question generation (fallback mode)
# ---------------------------------------------------------------------------
class QuestionGenerator:
"""
Generates candidate questions from KG triples using templates.
Fallback mode when TimelineKGQA is not available.
"""
def __init__(self, indexer: TemporalKGIndexer, seed: int = 42):
self.indexer = indexer
self.rng = random.Random(seed)
self.triples = indexer._all_triples
self.candidate_id_counter = 0
# Collect all actual timestamps for point-in-time TKGs (t_start == t_end)
self._timestamps: List[float] = sorted(
{t.t_start for t in self.triples}
)
# tkgl-smallpedia is 100% point-in-time; multi-hop chains need every
# hop at exactly the same year. Pre-index (subject, year) → triples
# so the 2-hop and 3+-hop samplers can forward-chain in O(bucket_size)
# instead of O(entity_degree across all years).
self._subject_year: Dict[str, Dict[float, List[Triple]]] = defaultdict(
lambda: defaultdict(list)
)
for t in self.triples:
self._subject_year[t.subject][t.t_start].append(t)
self._subjects_with_outgoing: List[str] = list(self._subject_year.keys())
# --- Public entry point ---------------------------------------------
def generate_candidates(
self,
target_candidates: int = 40000,
min_depth: int = 1,
max_depth: int = 4,
) -> List[Candidate]:
"""
Generate candidates across the 12-cell matrix (3 complexity levels ×
4 temporal operator types). Target per cell is an even split of the
per-complexity budget (4K/4K/2K per the TempBench paper).
Implementation: sample chains per complexity, and for each chain try
each of the four operator builders. PIT and INT succeed for almost
every chain; BA requires a sibling fact and SEQ requires an earlier
reference triple, so those buckets yield less per chain.
"""
# The TempBench paper commits to 4000 / 4000 / 2000 at target_n = 10000.
# Scale proportionally to target_candidates (= 4 × target_n from the
# BenchmarkBuilder).
per_complexity = {
"1hop": target_candidates // 10, # 40% of target_n
"2hop": target_candidates // 10, # 40% of target_n
"3plus": target_candidates // 20, # 20% of target_n
}
# 3× oversample to absorb Stage 3/4 attrition (uniqueness + dedup).
pool_per_complexity = {k: v * 3 for k, v in per_complexity.items()}
operators = ["point_in_time", "before_after", "interval", "sequence"]
candidates: List[Candidate] = []
for complexity, pool_target in pool_per_complexity.items():
sampler = {
"1hop": lambda n: self._sample_1hop_chains(n),
"2hop": lambda n: self._sample_2hop_chains(n),
"3plus": lambda n: self._sample_3plus_chains(n, max_depth),
}[complexity]
# Each chain can produce up to len(operators) candidates, so we
# only need pool_target / len(operators) chains — but many
# BA/SEQ attempts fail, so keep a healthy buffer.
chains = sampler(pool_target // 2)
for chain, t_query in chains:
for op in operators:
cand = self._build_operator_candidate(
op, chain, t_query, complexity
)
if cand is not None:
candidates.append(cand)
self.rng.shuffle(candidates)
return candidates
# --- Chain samplers (return chain + t_query, no templating) ---------
def _sample_1hop_chains(
self, target_count: int
) -> List[Tuple[List[Triple], float]]:
chains = []
sampled = self.rng.sample(
self.triples, min(len(self.triples), target_count)
)
for t in sampled:
chains.append(([t], t.t_start))
return chains
def _sample_2hop_chains(
self, target_count: int
) -> List[Tuple[List[Triple], float]]:
"""
Year-first 2-hop sampler. Pick an anchor subject that has outgoing
facts, pick a year from that subject, then find a second hop from the
pivot at the same year. Much higher success rate than sampling t1
first and hoping for a matching t2.
"""
chains = []
attempts = 0
max_attempts = target_count * 8
while len(chains) < target_count and attempts < max_attempts:
attempts += 1
anchor = self.rng.choice(self._subjects_with_outgoing)
year_index = self._subject_year[anchor]
year = self.rng.choice(list(year_index.keys()))
t1_candidates = year_index[year]
t1 = self.rng.choice(t1_candidates)
pivot_year_index = self._subject_year.get(t1.obj)
if pivot_year_index is None:
continue
t2_candidates = [
t for t in pivot_year_index.get(year, [])
if t.relation != t1.relation
and t.obj != t1.subject # cycle guard
]
if not t2_candidates:
continue
t2 = self.rng.choice(t2_candidates)
chains.append(([t1, t2], year))
return chains
def _sample_3plus_chains(
self, target_count: int, max_depth: int
) -> List[Tuple[List[Triple], float]]:
"""
Year-first 3+-hop sampler. Pick an anchor subject, pick a year that
subject has outgoing facts, walk forward through hops at that year.
Uses the pre-built (subject, year) index for fast forward-chaining.
"""
chains = []
attempts = 0
max_attempts = target_count * 15
while len(chains) < target_count and attempts < max_attempts:
attempts += 1
anchor = self.rng.choice(self._subjects_with_outgoing)
year_index = self._subject_year[anchor]
year = self.rng.choice(list(year_index.keys()))
depth = self.rng.randint(3, min(max_depth, 4))
chain: List[Triple] = []
current = anchor
visited = {current}
for _ in range(depth):
hops = [
t for t in self._subject_year.get(current, {}).get(year, [])
if t.obj not in visited
]
if not hops:
break
hop = self.rng.choice(hops)
chain.append(hop)
current = hop.obj
visited.add(current)
if len(chain) < 3:
continue
if any(
chain[i].relation == chain[i - 1].relation
for i in range(1, len(chain))
):
continue
window = ValidityWindow.full()
valid = True
for triple in chain:
window = TemporalKGIndexer.compose(window, triple, year)
if window is None or window.is_empty():
valid = False
break
if not valid:
continue
chains.append((chain, year))
return chains
# --- Operator builders ----------------------------------------------
def _build_operator_candidate(
self,
operator: str,
chain: List[Triple],
t_query: float,
complexity: str,
) -> Optional[Candidate]:
"""Dispatch to the right operator builder."""
dispatch = {
"point_in_time": self._build_pit,
"before_after": self._build_ba,
"interval": self._build_interval,
"sequence": self._build_sequence,
}
return dispatch[operator](chain, t_query, complexity)
def _build_pit(
self, chain: List[Triple], t_query: float, complexity: str
) -> Optional[Candidate]:
"""Point-in-time: "In {year}, what was {path}'s {r_n}?"."""
path = _render_composition_path(
chain[0].subject, [h.relation for h in chain[:-1]]
)
template = self.rng.choice(QUESTION_TEMPLATES["point_in_time"])
question = template.format(
path=path,
final_relation=chain[-1].relation,
year=int(t_query),
)
return self._make_candidate(
question, t_query, chain[-1].obj, chain, complexity,
"point_in_time", "explicit",
)
def _build_ba(
self, chain: List[Triple], t_query: float, complexity: str
) -> Optional[Candidate]:
"""
Before/after: same chain as PIT, but the temporal anchor is a sibling
fact (another holder of the terminal relation at a different time).
The candidate is dropped if no such sibling exists.
"""
terminal = chain[-1]
siblings = [
t for t in self.indexer.entity_index.get(terminal.subject, [])
if t.subject == terminal.subject
and t.relation == terminal.relation
and t.obj != terminal.obj
and t.t_start != t_query
]
if not siblings:
return None
sibling = self.rng.choice(siblings)
# Qualifier: answer came *before* sibling if t_query < sibling.t_start.
if sibling.t_start > t_query:
qualifier, qualifier_cap = "before", "Before"
else:
qualifier, qualifier_cap = "after", "After"
path = _render_composition_path(
chain[0].subject, [h.relation for h in chain[:-1]]
)
template = self.rng.choice(QUESTION_TEMPLATES["before_after"])
question = template.format(
path=path,
final_relation=terminal.relation,
qualifier=qualifier,
qualifier_cap=qualifier_cap,
sibling=sibling.obj,
)
return self._make_candidate(
question, t_query, terminal.obj, chain, complexity,
"before_after", "ordinal",
)
def _build_interval(
self, chain: List[Triple], t_query: float, complexity: str
) -> Optional[Candidate]:
"""
Interval (temporal pinpoint): "In what year was {path}'s {r_n} equal
to {terminal}?". Answer is the query year. Uniqueness of the year
is enforced by a custom check in Stage 3 (AnswerUniquenessFilter).
"""
path = _render_composition_path(
chain[0].subject, [h.relation for h in chain[:-1]]
)
terminal = chain[-1]
template = self.rng.choice(QUESTION_TEMPLATES["interval"])
question = template.format(
path=path,
final_relation=terminal.relation,
terminal=terminal.obj,
)
return self._make_candidate(
question, t_query, str(int(t_query)), chain, complexity,
"interval", "ordinal",
)
def _build_sequence(
self, chain: List[Triple], t_query: float, complexity: str
) -> Optional[Candidate]:
"""
Sequence: "After {ref_subject} {ref_rel} {ref_obj}, what was {path}'s
{r_n}?". Reference triple is an earlier fact involving any entity in
the chain (makes the anchor event contextually relevant).
"""
chain_entities = {h.subject for h in chain} | {h.obj for h in chain}
chain_ids = {(h.subject, h.relation, h.obj) for h in chain}
answer_entity = chain[-1].obj
refs = []
for e in chain_entities:
for t in self.indexer.entity_index.get(e, []):
if (
t.t_start < t_query
and (t.subject, t.relation, t.obj) not in chain_ids
# The reference must not contain the answer anywhere —
# otherwise the question literally names its own answer.
and t.subject != answer_entity
and t.obj != answer_entity
):
refs.append(t)
if not refs:
return None
ref = self.rng.choice(refs)
path = _render_composition_path(
chain[0].subject, [h.relation for h in chain[:-1]]
)
template = self.rng.choice(QUESTION_TEMPLATES["sequence"])
question = template.format(
path=path,
final_relation=chain[-1].relation,
ref_subject=ref.subject,
ref_relation=ref.relation,
ref_object=ref.obj,
)
return self._make_candidate(
question, t_query, chain[-1].obj, chain, complexity,
"sequence", "ordinal",
)
def _make_candidate(
self,
question: str,
t_query: float,
answer: str,
chain: List[Triple],
complexity: str,
operator_type: str,
question_type: str,
) -> Candidate:
cand = Candidate(
id=f"cand_{self.candidate_id_counter}",
question=question,
t_query=t_query,
answer=answer,
complexity=complexity,
operator_type=operator_type,
question_type=question_type,
gold_chain=chain,
)
self.candidate_id_counter += 1
return cand
# ---------------------------------------------------------------------------
# Stage 2: Composability filter
# ---------------------------------------------------------------------------
class ComposabilityFilter:
"""Validates temporal consistency of candidate chains."""
@staticmethod
def filter_candidates(candidates: List[Candidate]) -> List[Candidate]:
"""
Discard candidates whose gold chain has empty composed validity window.
Uses the ⊕ operator from ValidityWindow.intersect().
"""
valid = []
for cand in candidates:
# Compose all triples in the chain
window = ValidityWindow.full()
is_valid = True
for triple in cand.gold_chain:
window = TemporalKGIndexer.compose(window, triple, cand.t_query)
if window is None or window.is_empty():
is_valid = False
break
if is_valid:
valid.append(cand)
return valid
# ---------------------------------------------------------------------------
# Stage 3: Answer uniqueness filter
# ---------------------------------------------------------------------------
class AnswerUniquenessFilter:
"""Discards ambiguous questions with multiple valid answers at t_query."""
def __init__(self, indexer: TemporalKGIndexer):
self.indexer = indexer
def filter_candidates(self, candidates: List[Candidate]) -> List[Candidate]:
valid = [c for c in candidates if self._answer_is_unique(c)]
return valid
def _answer_is_unique(self, cand: Candidate) -> bool:
"""
Construction pipeline, Step 4: "removing ambiguous questions with multiple valid
answers at t_q". Scope is terminal-answer only — intermediate fan-out
along the chain is permitted.
Entity-valued operators (PIT, BA, SEQ): the final hop's (subject,
relation) must have exactly one valid object at t_q, and that object
must equal the recorded answer.
Year-valued operator (INT): the specific terminal triple (subject,
relation, object) must be valid at exactly one distinct t in the KG.
Otherwise the year-answer is ambiguous.
See benchmark-design-decisions.md §3 for the full rationale.
"""
if not cand.gold_chain:
return False
terminal = cand.gold_chain[-1]
if cand.operator_type == "interval":
# Year answer: the specific (s, r, o) triple must be unique in time.
matches = [
t for t in self.indexer.entity_index.get(terminal.subject, [])
if t.subject == terminal.subject
and t.relation == terminal.relation
and t.obj == terminal.obj
]
distinct_years = {t.t_start for t in matches}
return len(distinct_years) == 1
# Entity-valued operators: terminal (s, r) must have a single valid
# object at t_q.
terminal_objs = {
t.obj
for t in self.indexer.entity_index.get(terminal.subject, [])
if t.subject == terminal.subject
and t.relation == terminal.relation
and t.valid_at(cand.t_query)
}
return len(terminal_objs) == 1 and cand.answer in terminal_objs
# ---------------------------------------------------------------------------
# Stage 4: MinHash deduplication
# ---------------------------------------------------------------------------
class MinHasher:
"""
MinHash with Jaccard similarity for deduplicating near-duplicate questions.
Uses character 3-grams and 128 hash functions.
"""
def __init__(self, num_hashes: int = 128, gram_size: int = 3, seed: int = 42):
self.num_hashes = num_hashes
self.gram_size = gram_size
self.seed = seed
@staticmethod
def _get_grams(text: str, gram_size: int) -> Set[str]:
"""Extract character n-grams from text."""
text = text.lower()
return {text[i : i + gram_size] for i in range(len(text) - gram_size + 1)}
def _hash_gram(self, gram: str, hash_idx: int) -> int:
"""Compute hash value for a gram and hash function index."""
seed_str = f"{self.seed}_{hash_idx}_{gram}"
h = hashlib.sha256(seed_str.encode()).hexdigest()
return int(h, 16)
def signature(self, text: str) -> List[int]:
"""Compute MinHash signature for a text."""
grams = self._get_grams(text, self.gram_size)
if not grams:
return [0] * self.num_hashes
sig = []
for hash_idx in range(self.num_hashes):
min_hash = float("inf")
for gram in grams:
h = self._hash_gram(gram, hash_idx)
min_hash = min(min_hash, h)
sig.append(min_hash)
return sig
@staticmethod
def jaccard_similarity(sig_a: List[int], sig_b: List[int]) -> float:
"""Estimate Jaccard similarity from MinHash signatures."""
if len(sig_a) == 0 or len(sig_b) == 0:
return 0.0
matches = sum(1 for a, b in zip(sig_a, sig_b) if a == b)
return matches / len(sig_a)
class MinHashDeduplicator:
"""Deduplicates candidates using MinHash with Jaccard threshold."""
def __init__(self, similarity_threshold: float = 0.8, seed: int = 42):
self.threshold = similarity_threshold
self.hasher = MinHasher(num_hashes=128, seed=seed)
def deduplicate(self, candidates: List[Candidate]) -> List[Candidate]:
"""
Remove near-duplicate questions (Jaccard > threshold).
Keep the first occurrence of each group.
"""
if not candidates:
return []
# Compute signatures
sigs = [(c, self.hasher.signature(c.question)) for c in candidates]
# Greedy clustering: each candidate either starts a cluster or is merged
clusters = []
used = set()
for i, (cand_i, sig_i) in enumerate(sigs):
if i in used:
continue
cluster = [cand_i]
used.add(i)
for j in range(i + 1, len(sigs)):
if j in used:
continue
cand_j, sig_j = sigs[j]
sim = self.hasher.jaccard_similarity(sig_i, sig_j)
if sim > self.threshold:
used.add(j)
clusters.append(cluster[0]) # Keep first of each cluster
return clusters
# ---------------------------------------------------------------------------
# Stage 5: Subgraph construction
# ---------------------------------------------------------------------------
class SubgraphConstructor:
"""
Builds three subgraphs for each question:
- S_star: ground truth chain
- S_dist: distractor (one wrong hop, same relation, different object)
- S_stale: stale-fact (one hop replaced by temporally adjacent version)
"""
def __init__(self, indexer: TemporalKGIndexer, seed: int = 42):
self.indexer = indexer
self.rng = random.Random(seed)
def construct(
self,
candidate: Candidate,
) -> Tuple[List[Dict], List[Dict], List[Dict]]:
"""
Build S_star, S_dist, S_stale for a candidate.
Returns (S_star, S_dist, S_stale) as list of triple dicts.
"""
# S_star: the gold chain
S_star = [t.to_dict() for t in candidate.gold_chain]
# S_dist: replace one hop with a distractor (same relation, different object, valid at t_query)
S_dist = self._build_distractor(candidate)
# S_stale: replace one hop with a temporally adjacent version
S_stale = self._build_stale(candidate)
return S_star, S_dist, S_stale
def _build_distractor(self, candidate: Candidate) -> List[Dict]:
"""
Replace one random hop with a semantically similar but wrong triple.
Same subject, same relation, different object, valid at t_query.
"""
if not candidate.gold_chain:
return [t.to_dict() for t in candidate.gold_chain]
S_dist = [t.to_dict() for t in candidate.gold_chain]
hop_to_replace = self.rng.randint(0, len(candidate.gold_chain) - 1)
replaced_triple = candidate.gold_chain[hop_to_replace]
# Find alternative triples with same (s, r) but different object AT ANY TIME.
# We can't require valid_at(t_query) — the answer-uniqueness filter has
# already guaranteed no such alternatives exist, so that constraint would
# always return an empty list. Taking any alternative object across time
# gives a semantically plausible but factually wrong distractor (e.g. a
# former CEO when the question asks about a later year).
alternatives = [
t for t in self.indexer.entity_index.get(replaced_triple.subject, [])
if (t.subject == replaced_triple.subject and
t.relation == replaced_triple.relation and
t.obj != replaced_triple.obj)
]
if alternatives:
distractor = self.rng.choice(alternatives)
S_dist[hop_to_replace] = distractor.to_dict()
return S_dist
def _build_stale(self, candidate: Candidate) -> List[Dict]:
"""
Replace one hop with a temporally adjacent version
(same s, r, o but different validity window that does NOT contain t_query).
"""
if not candidate.gold_chain:
return [t.to_dict() for t in candidate.gold_chain]
S_stale = [t.to_dict() for t in candidate.gold_chain]
hop_to_replace = self.rng.randint(0, len(candidate.gold_chain) - 1)
replaced_triple = candidate.gold_chain[hop_to_replace]
# Find temporally adjacent versions of the same triple
adjacent = [
t for t in self.indexer._all_triples
if (t.subject == replaced_triple.subject and
t.relation == replaced_triple.relation and
t.obj == replaced_triple.obj and
t.t_start != replaced_triple.t_start and
not t.valid_at(candidate.t_query))
]
if adjacent:
stale_triple = self.rng.choice(adjacent)
S_stale[hop_to_replace] = stale_triple.to_dict()
return S_stale
# ---------------------------------------------------------------------------
# Stage 6: Train/dev/test split
# ---------------------------------------------------------------------------
class DatasetSplitter:
"""Stratified split by complexity level."""
def __init__(self, seed: int = 42):
self.rng = random.Random(seed)
def split(
self,
questions: List[BenchmarkQuestion],
train_frac: float = 0.7,
dev_frac: float = 0.1,
test_frac: float = 0.2,
) -> List[BenchmarkQuestion]:
"""
Stratified split by complexity.
Target: 1-hop: 4K (2.8K train), 2-hop: 4K (2.8K train), 3+-hop: 2K (1.4K train)
"""
# Group by complexity
by_complexity: Dict[str, List[BenchmarkQuestion]] = defaultdict(list)
for q in questions:
by_complexity[q.complexity].append(q)
# Split each group
result = []
for complexity, qs in by_complexity.items():
self.rng.shuffle(qs)
n = len(qs)
train_end = int(n * train_frac)
dev_end = train_end + int(n * dev_frac)
for i, q in enumerate(qs):
if i < train_end:
q.split = "train"
elif i < dev_end:
q.split = "dev"
else:
q.split = "test"
result.extend(qs)
return result
# ---------------------------------------------------------------------------
# Main builder
# ---------------------------------------------------------------------------
class BenchmarkBuilder:
"""Orchestrates all six pipeline stages."""
def __init__(
self,
kg_path: str,
output_dir: str = "./benchmark/",
target_n: int = 10000,
seed: int = 42,
min_depth: int = 1,
max_depth: int = 4,
):
self.kg_path = kg_path
self.output_dir = Path(output_dir)
self.target_n = target_n
self.seed = seed
self.min_depth = min_depth
self.max_depth = max_depth
self.rng = random.Random(seed)
# Build or load indexer
self.indexer = TemporalKGIndexer()
if kg_path.endswith(".json"):
self.indexer.load_from_json(kg_path)
else:
self.indexer.load_tkgl_smallpedia(kg_path)
def build(self) -> List[BenchmarkQuestion]:
"""Run the full 6-stage pipeline."""
print("[Benchmark Builder] Starting pipeline...")
print(f"[Indexer] {self.indexer}")
# Stage 1: Generate candidates
print("\n[Stage 1] Question generation...")
gen = QuestionGenerator(self.indexer, seed=self.seed)
target_candidates = int(self.target_n * 4) # 40K to yield 10K
candidates = gen.generate_candidates(
target_candidates=target_candidates,
min_depth=self.min_depth,
max_depth=self.max_depth,
)
print(f" Generated {len(candidates):,} candidates")
# Stage 2: Composability filter
print("\n[Stage 2] Composability filtering...")
comp_filter = ComposabilityFilter()
candidates = comp_filter.filter_candidates(candidates)
print(f" Passed composability check: {len(candidates):,}")
# Stage 3: Answer uniqueness filter
print("\n[Stage 3] Answer uniqueness filtering...")
uniq_filter = AnswerUniquenessFilter(self.indexer)
candidates = uniq_filter.filter_candidates(candidates)
print(f" Passed uniqueness check: {len(candidates):,}")
# Stage 4: MinHash deduplication
print("\n[Stage 4] MinHash deduplication...")
deduplicator = MinHashDeduplicator(similarity_threshold=0.8, seed=self.seed)
candidates = deduplicator.deduplicate(candidates)
print(f" After deduplication: {len(candidates):,}")
# Stratified trim: hit the paper's per-complexity targets
# (4K / 4K / 2K at target_n = 10K) and, within each complexity, spread
# across the four operator types. Under-yield in an operator cell
# is topped up with point-in-time from the same complexity, because
# PIT is the most reliable operator and degrades the benchmark least
# if it fills in for a short cell.
candidates = self._stratified_trim(candidates)
print(f" Trimmed to target: {len(candidates):,}")
# Stage 5: Subgraph construction
print("\n[Stage 5] Subgraph construction...")
sg_constructor = SubgraphConstructor(self.indexer, seed=self.seed)
questions = []
for i, cand in enumerate(candidates):
S_star, S_dist, S_stale = sg_constructor.construct(cand)
q = BenchmarkQuestion(
id=f"q_{i:06d}",
question=cand.question,
t_query=cand.t_query,
answer=cand.answer,
complexity=cand.complexity,
operator_type=cand.operator_type,
question_type=cand.question_type,
S_star=S_star,
S_dist=S_dist,
S_stale=S_stale,
split="", # Will be set in Stage 6
)
questions.append(q)
print(f" Subgraphs constructed: {len(questions):,}")
# Stage 6: Train/dev/test split
print("\n[Stage 6] Train/dev/test split (stratified)...")
splitter = DatasetSplitter(seed=self.seed)
questions = splitter.split(questions)
# Print split statistics
splits = defaultdict(lambda: defaultdict(int))
for q in questions:
splits[q.complexity][q.split] += 1
print(" Split distribution:")
for complexity in ["1hop", "2hop", "3plus"]:
if complexity in splits:
train = splits[complexity]["train"]
dev = splits[complexity]["dev"]
test = splits[complexity]["test"]
total = train + dev + test
print(f" {complexity:8s}: {total:5d} ({train:4d} train, {dev:3d} dev, {test:3d} test)")
return questions
def _stratified_trim(self, candidates: List[Candidate]) -> List[Candidate]:
"""
Reduce candidates to per-complexity/per-operator quotas matching paper
the TempBench paper (4K 1-hop, 4K 2-hop, 2K 3+-hop at target_n = 10K), with each
complexity split evenly across four operators. Cells that under-yield
are topped up from PIT in the same complexity.
"""
per_complexity_target = {
"1hop": self.target_n * 4 // 10, # 40%
"2hop": self.target_n * 4 // 10, # 40%
"3plus": self.target_n * 2 // 10, # 20%
}
operators = ("point_in_time", "before_after", "interval", "sequence")
# Bucket candidates by (complexity, operator)
buckets: Dict[Tuple[str, str], List[Candidate]] = defaultdict(list)
for c in candidates:
buckets[(c.complexity, c.operator_type)].append(c)
for bucket in buckets.values():
self.rng.shuffle(bucket)
trimmed: List[Candidate] = []
for complexity, cx_target in per_complexity_target.items():
per_op_target = cx_target // len(operators)
cx_kept: List[Candidate] = []
# First pass: take up to per_op_target from each operator cell.
for op in operators:
cell = buckets.get((complexity, op), [])
cx_kept.extend(cell[:per_op_target])
# Top up shortfall with PIT first, then anything available.
shortfall = cx_target - len(cx_kept)
if shortfall > 0:
topup_pool: List[Candidate] = []
for op in ("point_in_time", "sequence", "before_after", "interval"):
cell = buckets.get((complexity, op), [])
if len(cell) > per_op_target:
topup_pool.extend(cell[per_op_target:])
self.rng.shuffle(topup_pool)
cx_kept.extend(topup_pool[:shortfall])
trimmed.extend(cx_kept)
self.rng.shuffle(trimmed)
return trimmed
def save(self, questions: List[BenchmarkQuestion]) -> None:
"""Write benchmark to JSONL file."""
self.output_dir.mkdir(parents=True, exist_ok=True)
output_path = self.output_dir / "benchmark.jsonl"
with open(output_path, "w", encoding="utf-8") as f:
for q in questions:
f.write(json.dumps(q.to_jsonl_dict()) + "\n")
print(f"\n[Output] Benchmark saved to {output_path}")
print(f" {len(questions):,} questions written")
# ---------------------------------------------------------------------------
# Smoke test
# ---------------------------------------------------------------------------
def smoke_test():
"""Run on synthetic Deutsche Bank KG (same 5 triples as indexer.py tests)."""
print("=" * 70)
print("SMOKE TEST: TempBench Builder")
print("=" * 70)
# Create a temporary indexer with synthetic data
indexer = TemporalKGIndexer()
indexer.build([
("Deutsche_Bank", "has_CFO", "John_Cryan", 2015.0, 2018.0),
("Deutsche_Bank", "has_CFO", "Christian_Sewing", 2018.0, math.inf),
("Deutsche_Bank", "settled", "LIBOR_Case", 2015.25, 2015.25),
("John_Cryan", "member_of", "Deutsche_Bank", 2015.0, 2018.0),
("Christian_Sewing", "member_of", "Deutsche_Bank", 2018.0, math.inf),
])
print(f"\nIndexer: {indexer}\n")
# Generate candidates
gen = QuestionGenerator(indexer, seed=42)
candidates = gen.generate_candidates(
target_candidates=100,
min_depth=1,
max_depth=3,
)
print(f"Generated {len(candidates)} candidates")
# Filter by composability
comp_filter = ComposabilityFilter()
candidates = comp_filter.filter_candidates(candidates)
print(f"Passed composability: {len(candidates)}")
# Filter by uniqueness
uniq_filter = AnswerUniquenessFilter(indexer)
candidates = uniq_filter.filter_candidates(candidates)
print(f"Passed uniqueness: {len(candidates)}")
# Deduplicate
deduplicator = MinHashDeduplicator(similarity_threshold=0.8, seed=42)
candidates = deduplicator.deduplicate(candidates)
print(f"After deduplication: {len(candidates)}")
# Build subgraphs
sg_constructor = SubgraphConstructor(indexer, seed=42)
questions = []
for i, cand in enumerate(candidates[:min(5, len(candidates))]):
S_star, S_dist, S_stale = sg_constructor.construct(cand)
q = BenchmarkQuestion(
id=f"q_{i:06d}",
question=cand.question,
t_query=cand.t_query,
answer=cand.answer,
complexity=cand.complexity,
operator_type=cand.operator_type,
question_type=cand.question_type,
S_star=S_star,
S_dist=S_dist,
S_stale=S_stale,
split="train",
)
questions.append(q)
print(f"Built {len(questions)} benchmark questions\n")
# Print sample
print("Sample questions:")
for q in questions[:3]:
print(f"\n ID: {q.id}")
print(f" Question: {q.question}")
print(f" Complexity: {q.complexity}")
print(f" Answer: {q.answer}")
print(f" t_query: {q.t_query}")
print(f" Operator: {q.operator_type}")
print(f" S_star: {len(q.S_star)} triples")
print(f" S_dist: {len(q.S_dist)} triples")
print(f" S_stale: {len(q.S_stale)} triples")
print("\n" + "=" * 70)
print("SMOKE TEST PASSED")
print("=" * 70)
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def main():
parser = argparse.ArgumentParser(
description="Build TempBench: 10K-question temporal QA benchmark"
)
parser.add_argument(
"--kg_path",
type=str,
default=None,
help="Path to KG file (JSON or TKGL CSV)",
)
parser.add_argument(
"--output_dir",
type=str,
default="./benchmark/",
help="Output directory for benchmark JSONL",
)
parser.add_argument(
"--target_n",
type=int,
default=10000,
help="Target number of benchmark questions (default 10000)",
)
parser.add_argument(
"--seed",
type=int,
default=42,
help="Random seed (default 42)",
)
parser.add_argument(
"--min_depth",
type=int,
default=1,
help="Minimum chain depth (default 1)",
)
parser.add_argument(
"--max_depth",
type=int,
default=4,
help="Maximum chain depth (default 4)",
)
parser.add_argument(
"--smoke_test",
action="store_true",
help="Run smoke test on synthetic data",
)
args = parser.parse_args()
if args.smoke_test:
smoke_test()
sys.exit(0)
if not args.kg_path:
print("Error: --kg_path required (or use --smoke_test)")
sys.exit(1)
builder = BenchmarkBuilder(
kg_path=args.kg_path,
output_dir=args.output_dir,
target_n=args.target_n,
seed=args.seed,
min_depth=args.min_depth,
max_depth=args.max_depth,
)
questions = builder.build()
builder.save(questions)
if __name__ == "__main__":
main()