""" Knowledge Drift Dataset Statistics ==================================== Produces comprehensive statistics for the clean dataset. Outputs both console tables and a structured JSON report. Usage: python dataset_stats.py --dataset data/knowledge_drift_clean.json python dataset_stats.py --dataset data/knowledge_drift_clean.json --output data/stats/ """ import json import argparse import os import logging from collections import Counter, defaultdict from itertools import groupby logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') logger = logging.getLogger(__name__) def load_dataset(path): with open(path, 'r', encoding='utf-8') as f: return json.load(f) def print_section(title, width=90): print(f"\n{'='*width}") print(f" {title}") print(f"{'='*width}") def print_table(headers, rows, col_widths=None): """Print a formatted table.""" if col_widths is None: col_widths = [] for i, h in enumerate(headers): max_w = len(str(h)) for row in rows: if i < len(row): max_w = max(max_w, len(str(row[i]))) col_widths.append(max_w + 2) # Header header_line = "" for i, h in enumerate(headers): if i == 0: header_line += f" {str(h):<{col_widths[i]}}" else: header_line += f"{str(h):>{col_widths[i]}}" print(header_line) print(" " + "-" * (sum(col_widths) + len(col_widths))) # Rows for row in rows: line = "" for i, val in enumerate(row): if i == 0: line += f" {str(val):<{col_widths[i]}}" else: line += f"{str(val):>{col_widths[i]}}" print(line) def compute_stats(dataset): """Compute all statistics from the dataset.""" samples = dataset["samples"] metadata = dataset.get("metadata", {}) total = len(samples) stats = {"total_samples": total, "metadata": metadata} # ============================================================ # 1. CATEGORY DISTRIBUTION # ============================================================ cat_counts = Counter(s["category"] for s in samples) stats["categories"] = {k: {"count": v, "pct": round(100*v/total, 1)} for k, v in sorted(cat_counts.items())} # ============================================================ # 2. TEMPORAL ZONE DISTRIBUTION # ============================================================ zone_counts = Counter(s["temporal_zone"] for s in samples) stats["temporal_zones"] = {k: {"count": v, "pct": round(100*v/total, 1)} for k, v in sorted(zone_counts.items())} # ============================================================ # 3. CATEGORY × TEMPORAL ZONE CROSS-TAB # ============================================================ cross = defaultdict(lambda: defaultdict(int)) for s in samples: cross[s["category"]][s["temporal_zone"]] += 1 stats["category_x_zone"] = {cat: dict(zones) for cat, zones in sorted(cross.items())} # ============================================================ # 4. DRIFTED vs NON-DRIFTED BREAKDOWN # ============================================================ drifted = [s for s in samples if s.get("is_drifted_query")] not_drifted = [s for s in samples if not s.get("is_drifted_query")] post_cutoff = [s for s in samples if s["temporal_zone"] == "post_cutoff"] drifted_post = [s for s in drifted if s["temporal_zone"] == "post_cutoff"] not_drifted_post = [s for s in not_drifted if s["temporal_zone"] == "post_cutoff"] stats["drift_breakdown"] = { "total_drifted": len(drifted), "total_not_drifted": len(not_drifted), "post_cutoff_total": len(post_cutoff), "post_cutoff_drifted": len(drifted_post), "post_cutoff_not_drifted": len(not_drifted_post), "drift_ratio_post_cutoff": round(len(drifted_post) / max(len(post_cutoff), 1), 3), } # ============================================================ # 5. KNOWLEDGE TYPE DISTRIBUTION # ============================================================ kt_counts = Counter(s["knowledge_type"] for s in samples) stats["knowledge_types"] = {k: {"count": v, "pct": round(100*v/total, 1)} for k, v in sorted(kt_counts.items())} # Knowledge type × category kt_cat = defaultdict(lambda: defaultdict(int)) for s in samples: kt_cat[s["knowledge_type"]][s["category"]] += 1 stats["knowledge_type_x_category"] = {kt: dict(cats) for kt, cats in sorted(kt_cat.items())} # ============================================================ # 6. ENTITY ANALYSIS # ============================================================ all_entities = set(s["entity"] for s in samples) drifted_entities = set(s["entity"] for s in drifted) unchanged_entities = set(s["entity"] for s in samples if s["category"] == "no_drift" and s["knowledge_type"] == "entity_role") stable_entities = set(s["entity"] for s in samples if s["category"] == "stable") stats["entities"] = { "total_unique": len(all_entities), "drifted_unique": len(drifted_entities), "unchanged_unique": len(unchanged_entities), "stable_unique": len(stable_entities), } # Per-entity detail for drifted entity_detail = {} for s in drifted: ent = s["entity"] if ent not in entity_detail: entity_detail[ent] = { "relation": s.get("relation", ""), "old_answer": s.get("model_likely_answer", ""), "new_answer": s.get("expected_answer", ""), "change_date": s.get("change_date", ""), "knowledge_type": s.get("knowledge_type", ""), "num_queries": 0, } entity_detail[ent]["num_queries"] += 1 stats["drifted_entity_details"] = entity_detail # Per-entity detail for unchanged unchanged_detail = {} for s in samples: if s["category"] == "no_drift" and s["knowledge_type"] == "entity_role": ent = s["entity"] if ent not in unchanged_detail: unchanged_detail[ent] = { "relation": s.get("relation", ""), "answer": s.get("expected_answer", ""), "num_queries": 0, } unchanged_detail[ent]["num_queries"] += 1 stats["unchanged_entity_details"] = unchanged_detail # ============================================================ # 7. YEAR DISTRIBUTION # ============================================================ year_counts = Counter(s["year"] for s in samples) stats["year_distribution"] = {str(k): v for k, v in sorted(year_counts.items())} # Year × drift status year_drift = defaultdict(lambda: {"drifted": 0, "not_drifted": 0}) for s in samples: key = "drifted" if s.get("is_drifted_query") else "not_drifted" year_drift[s["year"]][key] += 1 stats["year_x_drift"] = {str(k): dict(v) for k, v in sorted(year_drift.items())} # ============================================================ # 8. TEMPLATE / QUERY ANALYSIS # ============================================================ # Count unique query templates (strip year) queries = [s["query"] for s in samples] unique_queries = len(set(queries)) # Identify question vs cloze format question_format = [s for s in samples if "?" in s["query"] or s["query"].lower().startswith("who") or s["query"].lower().startswith("what") or s["query"].lower().startswith("which")] cloze_format = [s for s in samples if "___" in s["query"] or s["query"].endswith(".")] other_format = [s for s in samples if s not in question_format and s not in cloze_format] stats["query_formats"] = { "unique_queries": unique_queries, "question_format": len(question_format), "cloze_format": len(cloze_format), "other_format": len(other_format), } # Question vs cloze × drifted q_drifted = len([s for s in question_format if s.get("is_drifted_query")]) c_drifted = len([s for s in cloze_format if s.get("is_drifted_query")]) stats["query_format_x_drift"] = { "question_drifted": q_drifted, "question_not_drifted": len(question_format) - q_drifted, "cloze_drifted": c_drifted, "cloze_not_drifted": len(cloze_format) - c_drifted, } # ============================================================ # 9. SOURCE DISTRIBUTION # ============================================================ src_counts = Counter(s.get("source", "unknown") for s in samples) stats["sources"] = dict(sorted(src_counts.items())) # ============================================================ # 10. CONFIDENCE DISTRIBUTION # ============================================================ conf_counts = Counter(s.get("confidence", "unknown") for s in samples) stats["confidence"] = dict(sorted(conf_counts.items())) return stats def print_report(stats): """Print a formatted report to console.""" print("\n" + "█" * 90) print("█" + " " * 30 + "DATASET STATISTICS REPORT" + " " * 34 + "█") print("█" * 90) # ---- OVERVIEW ---- print_section("1. OVERVIEW") print(f" Total samples: {stats['total_samples']}") print(f" Model: {stats['metadata'].get('model', 'N/A')}") print(f" Model cutoff: {stats['metadata'].get('model_cutoff', 'N/A')}") print(f" Unique entities: {stats['entities']['total_unique']}") print(f" Unique queries: {stats['query_formats']['unique_queries']}") # ---- CATEGORY DISTRIBUTION ---- print_section("2. CATEGORY DISTRIBUTION") rows = [] for cat, info in stats["categories"].items(): desc = { "stable": "Timeless facts (capitals, science, math)", "no_drift": "Could change but didn't (same leader/CEO)", "unknown_drift": "Changed after cutoff (model doesn't know)", }.get(cat, "") rows.append([cat, info["count"], f"{info['pct']}%", desc]) print_table(["Category", "Count", "%", "Description"], rows, [22, 8, 8, 48]) # ---- TEMPORAL ZONE ---- print_section("3. TEMPORAL ZONE DISTRIBUTION") rows = [[z, info["count"], f"{info['pct']}%"] for z, info in stats["temporal_zones"].items()] print_table(["Zone", "Count", "%"], rows, [18, 8, 8]) # ---- CROSS-TAB: Category × Zone ---- print_section("4. CATEGORY × TEMPORAL ZONE") zones = sorted(set(z for zones in stats["category_x_zone"].values() for z in zones)) headers = ["Category"] + zones + ["Total"] rows = [] for cat, zone_counts in sorted(stats["category_x_zone"].items()): row = [cat] + [zone_counts.get(z, 0) for z in zones] row.append(sum(zone_counts.values())) rows.append(row) print_table(headers, rows, [22] + [14]*len(zones) + [8]) # ---- DRIFT BREAKDOWN ---- print_section("5. DRIFT ANALYSIS (KEY EXPERIMENTAL GROUPS)") db = stats["drift_breakdown"] print(f"\n POST-CUTOFF SAMPLES (the experimental comparison):") print(f" Total post-cutoff: {db['post_cutoff_total']}") print(f" ├── DRIFTED: {db['post_cutoff_drifted']} ← answer changed, model doesn't know") print(f" └── NOT DRIFTED: {db['post_cutoff_not_drifted']} ← answer unchanged, model correct") print(f" Drift ratio: {db['drift_ratio_post_cutoff']:.1%}") print(f"\n ALL SAMPLES:") print(f" Drifted queries: {db['total_drifted']}") print(f" Non-drifted queries: {db['total_not_drifted']}") # ---- DRIFTED ENTITIES ---- print_section("6. DRIFTED ENTITIES (Post-Cutoff Changes)") rows = [] for ent, info in sorted(stats["drifted_entity_details"].items()): rows.append([ ent, info["relation"], info["old_answer"], info["new_answer"], info["change_date"], info["num_queries"] ]) print_table( ["Entity", "Relation", "Old (pre-cutoff)", "New (post-cutoff)", "Changed", "#Q"], rows, [18, 14, 28, 28, 12, 5] ) # ---- UNCHANGED ENTITIES ---- print_section("7. UNCHANGED ENTITIES (Same Type, No Drift — Controls)") rows = [] for ent, info in sorted(stats["unchanged_entity_details"].items()): rows.append([ent, info["relation"], info["answer"], info["num_queries"]]) print_table(["Entity", "Relation", "Answer (stable)", "#Q"], rows, [18, 16, 32, 5]) # ---- KNOWLEDGE TYPE ---- print_section("8. KNOWLEDGE TYPE DISTRIBUTION") rows = [] for kt, info in stats["knowledge_types"].items(): rows.append([kt, info["count"], f"{info['pct']}%"]) print_table(["Type", "Count", "%"], rows, [18, 8, 8]) # Knowledge type × category print("\n Knowledge Type × Category:") for kt, cats in sorted(stats["knowledge_type_x_category"].items()): parts = ", ".join(f"{c}={n}" for c, n in sorted(cats.items())) print(f" {kt:<18} {parts}") # ---- YEAR DISTRIBUTION ---- print_section("9. YEAR DISTRIBUTION") rows = [] for year, count in sorted(stats["year_distribution"].items()): drift_info = stats["year_x_drift"].get(year, {}) d = drift_info.get("drifted", 0) nd = drift_info.get("not_drifted", 0) rows.append([year, count, d, nd]) print_table(["Year", "Total", "Drifted", "Not Drifted"], rows, [8, 8, 10, 14]) # ---- QUERY FORMAT ---- print_section("10. QUERY FORMAT ANALYSIS") qf = stats["query_formats"] print(f" Question format (Who is...?): {qf['question_format']}") print(f" Cloze format (X was ___.): {qf['cloze_format']}") print(f" Other: {qf['other_format']}") qfd = stats["query_format_x_drift"] print(f"\n Format × Drift status:") print(f" Question — drifted: {qfd['question_drifted']}, not drifted: {qfd['question_not_drifted']}") print(f" Cloze — drifted: {qfd['cloze_drifted']}, not drifted: {qfd['cloze_not_drifted']}") # ---- DATA QUALITY ---- print_section("11. DATA QUALITY") src = stats["sources"] conf = stats["confidence"] print(f" Sources:") for s, c in src.items(): print(f" {s}: {c}") print(f" Confidence levels:") for c, n in conf.items(): print(f" {c}: {n}") # ---- PAPER-READY SUMMARY ---- print_section("12. PAPER-READY SUMMARY") db = stats["drift_breakdown"] ent = stats["entities"] print(f""" "We construct a temporally-scoped factual knowledge dataset comprising {stats['total_samples']} query instances across {ent['total_unique']} unique entities. The dataset spans three categories: - {stats['categories'].get('unknown_drift', {}).get('count', 0)} queries about {ent['drifted_unique']} entities whose facts changed after the model's knowledge cutoff (August 2024), - {stats['categories'].get('no_drift', {}).get('count', 0)} queries about {ent['unchanged_unique']} entities of the same type that did NOT change (controls), and - {stats['categories'].get('stable', {}).get('count', 0)} queries about timeless facts (geography, science, math). Each query is instantiated across {len(stats['year_distribution'])} time points ({', '.join(sorted(stats['year_distribution'].keys()))}), yielding {db['post_cutoff_drifted']} drifted and {db['post_cutoff_not_drifted']} non-drifted post-cutoff instances for the primary experimental comparison." """) print("=" * 90) def main(): parser = argparse.ArgumentParser(description="Dataset Statistics") parser.add_argument("--dataset", default="data/knowledge_drift_clean.json") parser.add_argument("--output", default=None, help="Output directory for JSON stats") args = parser.parse_args() dataset = load_dataset(args.dataset) stats = compute_stats(dataset) print_report(stats) # Save JSON if args.output: os.makedirs(args.output, exist_ok=True) out_path = os.path.join(args.output, "dataset_statistics.json") with open(out_path, 'w') as f: json.dump(stats, f, indent=2, ensure_ascii=False) logger.info(f"Statistics saved to {out_path}") if __name__ == "__main__": main()