""" Dataset Validation & Cleaning for Knowledge Drift Detection ============================================================= The Wikidata SPARQL queries have critical data quality issues: PROBLEM: SPARQL temporal qualifiers return ALL historical holders, not just the most recent change. So "head of state of Romania in 2025" might return "Nicolae Ceaușescu" (executed 1989) because his end_date is recorded as a datetime that Wikidata parsed oddly. This script: 1. Diagnoses all samples for obvious errors 2. Filters to ONLY manually verified + web-validated facts 3. Expands the verified set with additional curated examples 4. Produces a clean dataset ready for experiments Usage: python validate_and_clean_dataset.py \ --input data/knowledge_drift_dataset.json \ --output data/knowledge_drift_clean.json """ import json import argparse import logging import os from datetime import datetime from collections import Counter, defaultdict logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') logger = logging.getLogger(__name__) # ============================================================ # EXPANDED VERIFIED FACTS (Manually researched, high confidence) # ============================================================ # Each entry has: entity, relation, old_answer (pre-cutoff), new_answer (post-cutoff), # change_date, and query templates. These are the GOLD STANDARD. VERIFIED_DRIFTED_FACTS = [ # === POLITICAL LEADERS (highest confidence) === { "entity": "United Kingdom", "relation": "Prime Minister", "knowledge_type": "entity_role", "old_answer": "Rishi Sunak", "new_answer": "Keir Starmer", "change_date": "2024-07-05", "confidence": "high", "source": "public_record", "templates": [ "Who is the Prime Minister of the United Kingdom in {year}?", "In {year}, the Prime Minister of the UK was ___.", "As of {year}, who leads the British government?", ], }, { "entity": "United States", "relation": "President", "knowledge_type": "entity_role", "old_answer": "Joe Biden", "new_answer": "Donald Trump", "change_date": "2025-01-20", "confidence": "high", "source": "public_record", "templates": [ "Who is the President of the United States in {year}?", "In {year}, the US President was ___.", "As of {year}, who is the American president?", ], }, { "entity": "Japan", "relation": "Prime Minister", "knowledge_type": "entity_role", "old_answer": "Fumio Kishida", "new_answer": "Shigeru Ishiba", "change_date": "2024-10-01", "confidence": "high", "source": "public_record", "templates": [ "Who is the Prime Minister of Japan in {year}?", "In {year}, Japan's Prime Minister was ___.", ], }, { "entity": "Syria", "relation": "Leader", "knowledge_type": "entity_role", "old_answer": "Bashar al-Assad", "new_answer": "Ahmad al-Sharaa", "change_date": "2024-12-08", "confidence": "high", "source": "public_record", "templates": [ "Who is the leader of Syria in {year}?", "In {year}, Syria was led by ___.", "As of {year}, who controls Syria?", ], }, { "entity": "Mexico", "relation": "President", "knowledge_type": "entity_role", "old_answer": "Andrés Manuel López Obrador", "new_answer": "Claudia Sheinbaum", "change_date": "2024-10-01", "confidence": "high", "source": "public_record", "templates": [ "Who is the President of Mexico in {year}?", "In {year}, Mexico's President was ___.", ], }, { "entity": "Indonesia", "relation": "President", "knowledge_type": "entity_role", "old_answer": "Joko Widodo", "new_answer": "Prabowo Subianto", "change_date": "2024-10-20", "confidence": "high", "source": "public_record", "templates": [ "Who is the President of Indonesia in {year}?", "In {year}, Indonesia's President was ___.", ], }, { "entity": "India", "relation": "Prime Minister", "knowledge_type": "entity_role", "old_answer": "Narendra Modi", "new_answer": "Narendra Modi", # re-elected — CONTROL "change_date": "2024-06-09", "confidence": "high", "source": "public_record", "note": "CONTROL: same person re-elected, answer unchanged despite event", "templates": [ "Who is the Prime Minister of India in {year}?", "In {year}, India's Prime Minister was ___.", ], }, { "entity": "South Korea", "relation": "President", "knowledge_type": "entity_role", "old_answer": "Yoon Suk-yeol", "new_answer": "Han Duck-soo", "change_date": "2024-12-14", "confidence": "medium", "source": "public_record", "note": "Yoon impeached Dec 2024, acting president Han Duck-soo", "templates": [ "Who is the President of South Korea in {year}?", "In {year}, South Korea's President was ___.", ], }, { "entity": "Sri Lanka", "relation": "President", "knowledge_type": "entity_role", "old_answer": "Ranil Wickremesinghe", "new_answer": "Anura Kumara Dissanayake", "change_date": "2024-09-23", "confidence": "high", "source": "public_record", "templates": [ "Who is the President of Sri Lanka in {year}?", "In {year}, Sri Lanka's President was ___.", ], }, # === CORPORATE === { "entity": "Intel", "relation": "CEO", "knowledge_type": "entity_role", "old_answer": "Pat Gelsinger", "new_answer": "Lip-Bu Tan", "change_date": "2024-12-01", "confidence": "high", "source": "public_record", "note": "Gelsinger resigned Dec 2024; interim then Lip-Bu Tan appointed Mar 2025", "templates": [ "Who is the CEO of Intel in {year}?", "In {year}, Intel's CEO was ___.", ], }, { "entity": "Nike", "relation": "CEO", "knowledge_type": "entity_role", "old_answer": "John Donahoe", "new_answer": "Elliott Hill", "change_date": "2024-10-14", "confidence": "high", "source": "public_record", "templates": [ "Who is the CEO of Nike in {year}?", "In {year}, Nike's CEO was ___.", ], }, { "entity": "Starbucks", "relation": "CEO", "knowledge_type": "entity_role", "old_answer": "Laxman Narasimhan", "new_answer": "Brian Niccol", "change_date": "2024-09-09", "confidence": "high", "source": "public_record", "templates": [ "Who is the CEO of Starbucks in {year}?", "In {year}, the CEO of Starbucks was ___.", ], }, # === ORGANIZATIONS / GEOPOLITICAL === { "entity": "BRICS", "relation": "member countries", "knowledge_type": "relational", "old_answer": "Brazil, Russia, India, China, South Africa", "new_answer": "Brazil, Russia, India, China, South Africa, Iran, Egypt, Ethiopia, UAE, Saudi Arabia", "change_date": "2024-01-01", "confidence": "high", "source": "public_record", "note": "Expansion announced Aug 2023, effective Jan 2024. Near cutoff — model may partially know.", "templates": [ "Which countries are members of BRICS in {year}?", "In {year}, the BRICS members included ___.", ], }, { "entity": "OpenAI", "relation": "corporate structure", "knowledge_type": "relational", "old_answer": "nonprofit with capped-profit subsidiary", "new_answer": "public benefit corporation (for-profit transition)", "change_date": "2024-12-27", "confidence": "medium", "source": "public_record", "note": "Announced transition to PBC in late 2024", "templates": [ "What is OpenAI's corporate structure in {year}?", "In {year}, OpenAI was organized as ___.", ], }, ] # === VERIFIED UNCHANGED FACTS (Category 4: same type, didn't change) === VERIFIED_UNCHANGED_FACTS = [ {"entity": "Saudi Arabia", "relation": "King", "answer": "King Salman bin Abdulaziz", "since": "2015", "templates": ["Who is the King of Saudi Arabia in {year}?", "In {year}, the King of Saudi Arabia was ___."]}, {"entity": "UAE", "relation": "President", "answer": "Mohamed bin Zayed Al Nahyan", "since": "2022", "templates": ["Who is the President of the UAE in {year}?", "In {year}, the UAE President was ___."]}, {"entity": "Russia", "relation": "President", "answer": "Vladimir Putin", "since": "2012", "templates": ["Who is the President of Russia in {year}?", "In {year}, Russia's President was ___."]}, {"entity": "China", "relation": "President", "answer": "Xi Jinping", "since": "2013", "templates": ["Who is the President of China in {year}?", "In {year}, China's President was ___."]}, {"entity": "France", "relation": "President", "answer": "Emmanuel Macron", "since": "2017", "templates": ["Who is the President of France in {year}?", "In {year}, France's President was ___."]}, {"entity": "Egypt", "relation": "President", "answer": "Abdel Fattah el-Sisi", "since": "2014", "templates": ["Who is the President of Egypt in {year}?", "In {year}, Egypt's President was ___."]}, {"entity": "Apple", "relation": "CEO", "answer": "Tim Cook", "since": "2011", "templates": ["Who is the CEO of Apple in {year}?", "In {year}, Apple's CEO was ___."]}, {"entity": "Microsoft", "relation": "CEO", "answer": "Satya Nadella", "since": "2014", "templates": ["Who is the CEO of Microsoft in {year}?", "In {year}, Microsoft's CEO was ___."]}, {"entity": "Amazon", "relation": "CEO", "answer": "Andy Jassy", "since": "2021", "templates": ["Who is the CEO of Amazon in {year}?", "In {year}, Amazon's CEO was ___."]}, {"entity": "Tesla", "relation": "CEO", "answer": "Elon Musk", "since": "2008", "templates": ["Who is the CEO of Tesla in {year}?", "In {year}, Tesla's CEO was ___."]}, {"entity": "Google", "relation": "CEO", "answer": "Sundar Pichai", "since": "2015", "templates": ["Who is the CEO of Google in {year}?", "In {year}, Google's CEO was ___."]}, {"entity": "Turkey", "relation": "President", "answer": "Recep Tayyip Erdoğan", "since": "2014", "templates": ["Who is the President of Turkey in {year}?", "In {year}, Turkey's President was ___."]}, {"entity": "Germany", "relation": "Chancellor", "answer": "Olaf Scholz", "since": "2021", "templates": ["Who is the Chancellor of Germany in {year}?", "In {year}, Germany's Chancellor was ___."]}, {"entity": "Canada", "relation": "Prime Minister", "answer": "Justin Trudeau", "since": "2015", "templates": ["Who is the Prime Minister of Canada in {year}?", "In {year}, Canada's PM was ___."]}, {"entity": "Israel", "relation": "Prime Minister", "answer": "Benjamin Netanyahu", "since": "2022", "templates": ["Who is the Prime Minister of Israel in {year}?", "In {year}, Israel's PM was ___."]}, {"entity": "Meta", "relation": "CEO", "answer": "Mark Zuckerberg", "since": "2004", "templates": ["Who is the CEO of Meta in {year}?", "In {year}, Meta's CEO was ___."]}, {"entity": "NVIDIA", "relation": "CEO", "answer": "Jensen Huang", "since": "1993", "templates": ["Who is the CEO of NVIDIA in {year}?", "In {year}, NVIDIA's CEO was ___."]}, ] # === STABLE FACTS (Category 1: timeless, never change) === STABLE_FACTS_EXPANDED = [ # Geography {"query": "What is the capital of France?", "answer": "Paris", "type": "geographical"}, {"query": "What is the capital of Japan?", "answer": "Tokyo", "type": "geographical"}, {"query": "What is the capital of Egypt?", "answer": "Cairo", "type": "geographical"}, {"query": "What is the capital of Germany?", "answer": "Berlin", "type": "geographical"}, {"query": "What is the capital of Saudi Arabia?", "answer": "Riyadh", "type": "geographical"}, {"query": "What is the capital of UAE?", "answer": "Abu Dhabi", "type": "geographical"}, {"query": "What is the capital of Brazil?", "answer": "Brasília", "type": "geographical"}, {"query": "What is the capital of Australia?", "answer": "Canberra", "type": "geographical"}, {"query": "What is the capital of South Korea?", "answer": "Seoul", "type": "geographical"}, {"query": "What is the capital of Morocco?", "answer": "Rabat", "type": "geographical"}, {"query": "What is the longest river in the world?", "answer": "the Nile", "type": "geographical"}, {"query": "What is the highest mountain in the world?", "answer": "Mount Everest", "type": "geographical"}, {"query": "What is the largest ocean?", "answer": "the Pacific Ocean", "type": "geographical"}, # Science {"query": "What is the chemical formula for water?", "answer": "H2O", "type": "scientific"}, {"query": "What is the speed of light in km/s?", "answer": "approximately 300,000", "type": "scientific"}, {"query": "What does DNA stand for?", "answer": "deoxyribonucleic acid", "type": "scientific"}, {"query": "What is the force of gravity on Earth in m/s²?", "answer": "9.8", "type": "scientific"}, {"query": "What is absolute zero in Celsius?", "answer": "-273.15", "type": "scientific"}, # Math {"query": "What is the value of pi to two decimal places?", "answer": "3.14", "type": "mathematical"}, {"query": "What is the square root of 144?", "answer": "12", "type": "mathematical"}, {"query": "What is 2 to the power of 10?", "answer": "1024", "type": "mathematical"}, # History {"query": "When did World War II end?", "answer": "1945", "type": "historical"}, {"query": "When was the first moon landing?", "answer": "1969", "type": "historical"}, {"query": "When did the Berlin Wall fall?", "answer": "1989", "type": "historical"}, {"query": "Who painted the Mona Lisa?", "answer": "Leonardo da Vinci", "type": "historical"}, ] MODEL_CUTOFF = "2024-08-01" TIMESTAMP_YEARS = [2020, 2022, 2023, 2024, 2025] def diagnose_existing_dataset(dataset_path): """Analyze existing dataset for quality issues.""" with open(dataset_path, 'r') as f: dataset = json.load(f) samples = dataset.get("samples", []) logger.info(f"Loaded {len(samples)} samples from {dataset_path}") issues = [] stats = Counter() # Known dead people / historical figures that should never appear as 2025 answers impossible_2025 = [ "abraham lincoln", "Nicolae Ceaușescu".lower(), "george washington", "winston churchill", "joseph stalin", "mao zedong", "gandhi", "john f. kennedy", "josef klaus", "walter zenga", "wolfgang schüssel", ] for i, s in enumerate(samples): stats[s.get("category", "unknown")] += 1 stats[f"drifted={s.get('is_drifted_query', False)}"] += 1 stats[f"source={s.get('source', 'unknown')}"] += 1 # Check for obviously wrong answers exp = s.get("expected_answer", "").lower() old = s.get("model_likely_answer", "").lower() year = s.get("year", 0) entity = s.get("entity", "") # Flag: historical figure as 2025 answer for imp in impossible_2025: if imp in exp and year >= 2024: issues.append({ "index": i, "severity": "CRITICAL", "issue": f"Historical/dead figure '{s.get('expected_answer')}' as {year} answer for {entity}", "query": s.get("query", ""), "expected": s.get("expected_answer"), "source": s.get("source"), }) # Flag: old and new answer are the same (not a real drift) if s.get("is_drifted_query") and exp == old and exp: issues.append({ "index": i, "severity": "WARNING", "issue": f"Drifted query but old==new answer: '{exp}'", "query": s.get("query", ""), "entity": entity, }) # Flag: duplicate queries for same entity+year with different answers # (collected separately below) # Check for duplicate entity+year with conflicting answers entity_year_answers = defaultdict(list) for i, s in enumerate(samples): key = (s.get("entity", ""), s.get("year", 0), s.get("relation", "")) entity_year_answers[key].append({ "index": i, "expected": s.get("expected_answer", ""), "query": s.get("query", ""), }) for key, entries in entity_year_answers.items(): answers = set(e["expected"] for e in entries) if len(answers) > 1: issues.append({ "severity": "CRITICAL", "issue": f"CONFLICTING answers for {key}: {answers}", "entries": entries, }) # Print report print("\n" + "=" * 80) print(" DATASET QUALITY DIAGNOSIS") print("=" * 80) print(f"\nTotal samples: {len(samples)}") print("\nBy category:") for k, v in sorted(stats.items()): print(f" {k}: {v}") print(f"\n{'='*80}") print(f" ISSUES FOUND: {len(issues)}") print(f"{'='*80}") critical = [i for i in issues if i["severity"] == "CRITICAL"] warnings = [i for i in issues if i["severity"] == "WARNING"] print(f"\n CRITICAL: {len(critical)}") for issue in critical[:20]: print(f" ❌ {issue['issue']}") if 'query' in issue: print(f" Query: {issue.get('query', '')[:80]}") print(f"\n WARNINGS: {len(warnings)}") for issue in warnings[:10]: print(f" ⚠️ {issue['issue']}") return issues, stats def build_clean_dataset(): """Build a clean dataset from ONLY verified facts.""" samples = [] sample_id = 0 # === 1. DRIFTED FACTS (Category 3: unknown_drift) === for fact in VERIFIED_DRIFTED_FACTS: is_control = (fact["old_answer"] == fact["new_answer"]) for template in fact["templates"]: for year in TIMESTAMP_YEARS: change_year = int(fact["change_date"][:4]) change_month = int(fact["change_date"][5:7]) # Determine correct answer for this year if year < change_year: expected = fact["old_answer"] is_drifted = False elif year == change_year and change_month > 8: # Changed after Qwen cutoff (Aug 2024) in the same year # For 2024 queries, model likely knows old answer expected = fact["old_answer"] is_drifted = False else: expected = fact["new_answer"] is_drifted = not is_control and year > 2024 # Temporal zone if year < 2024: temporal_zone = "pre_cutoff" elif year == 2024: temporal_zone = "near_cutoff" else: temporal_zone = "post_cutoff" query = template.format(year=year) sample = { "id": f"verified_{sample_id:04d}", "query": query, "expected_answer": expected, "model_likely_answer": fact["old_answer"] if is_drifted else "", "entity": fact["entity"], "relation": fact["relation"], "knowledge_type": fact["knowledge_type"], "category": "unknown_drift" if not is_control else "no_drift", "is_drifted_query": is_drifted, "year": year, "temporal_zone": temporal_zone, "change_date": fact["change_date"], "confidence": fact.get("confidence", "high"), "source": "manual_verified", "note": fact.get("note", ""), } samples.append(sample) sample_id += 1 # === 2. UNCHANGED FACTS (Category 4: no_drift, same type) === for fact in VERIFIED_UNCHANGED_FACTS: for template in fact["templates"]: for year in TIMESTAMP_YEARS: if year < 2024: temporal_zone = "pre_cutoff" elif year == 2024: temporal_zone = "near_cutoff" else: temporal_zone = "post_cutoff" query = template.format(year=year) sample = { "id": f"unchanged_{sample_id:04d}", "query": query, "expected_answer": fact["answer"], "model_likely_answer": "", "entity": fact["entity"], "relation": fact["relation"], "knowledge_type": "entity_role", "category": "no_drift", "is_drifted_query": False, "year": year, "temporal_zone": temporal_zone, "confidence": "high", "source": "manual_verified", } samples.append(sample) sample_id += 1 # === 3. STABLE FACTS (Category 1: timeless) === for fact in STABLE_FACTS_EXPANDED: for year in TIMESTAMP_YEARS: if year < 2024: temporal_zone = "pre_cutoff" elif year == 2024: temporal_zone = "near_cutoff" else: temporal_zone = "post_cutoff" query = f"In {year}, {fact['query'].lower()}" if not fact["query"].startswith("In") else fact["query"] sample = { "id": f"stable_{sample_id:04d}", "query": query, "expected_answer": fact["answer"], "model_likely_answer": "", "entity": fact["query"].split("?")[0] if "?" in fact["query"] else fact["query"], "relation": "fact", "knowledge_type": fact["type"], "category": "stable", "is_drifted_query": False, "year": year, "temporal_zone": temporal_zone, "confidence": "high", "source": "manual_curated", } samples.append(sample) sample_id += 1 return samples def print_dataset_stats(samples): """Print comprehensive stats for the clean dataset.""" print("\n" + "=" * 80) print(" CLEAN DATASET STATISTICS") print("=" * 80) total = len(samples) print(f"\nTotal samples: {total}") # By category cats = Counter(s["category"] for s in samples) print("\nBy category:") for k, v in sorted(cats.items()): print(f" {k}: {v} ({100*v/total:.1f}%)") # By temporal zone zones = Counter(s["temporal_zone"] for s in samples) print("\nBy temporal zone:") for k, v in sorted(zones.items()): print(f" {k}: {v}") # Drifted queries drifted = [s for s in samples if s["is_drifted_query"]] not_drifted_post = [s for s in samples if not s["is_drifted_query"] and s["temporal_zone"] == "post_cutoff"] print(f"\nDrifted queries (post-cutoff, answer changed): {len(drifted)}") print(f"Non-drifted post-cutoff queries: {len(not_drifted_post)}") print(f"Ratio drifted / non-drifted (post-cutoff): {len(drifted)}/{len(not_drifted_post)}") # By entity (drifted only) if drifted: ent = Counter(s["entity"] for s in drifted) print(f"\nDrifted entities ({len(ent)} unique):") for k, v in ent.most_common(15): print(f" {k}: {v} queries") # By knowledge type kt = Counter(s["knowledge_type"] for s in samples) print("\nBy knowledge type:") for k, v in sorted(kt.items()): print(f" {k}: {v}") # By source src = Counter(s["source"] for s in samples) print("\nBy source:") for k, v in sorted(src.items()): print(f" {k}: {v}") def main(): parser = argparse.ArgumentParser() parser.add_argument("--input", default="data/knowledge_drift_dataset.json", help="Original dataset to diagnose") parser.add_argument("--output", default="data/knowledge_drift_clean.json", help="Clean dataset output path") parser.add_argument("--diagnose_only", action="store_true", help="Only diagnose, don't build clean dataset") args = parser.parse_args() # === Step 1: Diagnose existing dataset === if os.path.exists(args.input): print("\n" + "█" * 80) print(" DIAGNOSING EXISTING DATASET") print("█" * 80) issues, stats = diagnose_existing_dataset(args.input) else: logger.warning(f"No existing dataset at {args.input}, skipping diagnosis") if args.diagnose_only: return # === Step 2: Build clean dataset === print("\n" + "█" * 80) print(" BUILDING CLEAN DATASET") print("█" * 80) samples = build_clean_dataset() print_dataset_stats(samples) # === Step 3: Save === os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True) dataset = { "metadata": { "description": "Clean knowledge drift detection dataset (verified facts only)", "model": "Qwen/Qwen2.5-7B-Instruct", "model_cutoff": MODEL_CUTOFF, "created": datetime.now().isoformat(), "num_samples": len(samples), "num_drifted": len([s for s in samples if s["is_drifted_query"]]), "num_entities_drifted": len(set(s["entity"] for s in samples if s["is_drifted_query"])), "categories": dict(Counter(s["category"] for s in samples)), "note": "Built from manually verified facts only. No Wikidata SPARQL (unreliable temporal data).", }, "samples": samples, } with open(args.output, 'w', encoding='utf-8') as f: json.dump(dataset, f, indent=2, ensure_ascii=False) logger.info(f"\n✅ Clean dataset saved to {args.output}") logger.info(f" {len(samples)} samples, {dataset['metadata']['num_drifted']} drifted") logger.info(f"\n Run experiments with:") logger.info(f" python run_experiments.py --dataset {args.output} --all --max_samples 200") if __name__ == "__main__": main()