data_diagnostic.py

"""
Data Diagnostic: Answers all reviewer questions about the dataset.
Run: python data_diagnostic.py
"""
import json
from collections import Counter, defaultdict
import numpy as np

# Load both datasets
print("=" * 70)
print("  DATA DIAGNOSTIC — Answering All Reviewer Questions")
print("=" * 70)

with open("data/knowledge_drift_unified_tier1.json") as f:
    tier1 = json.load(f)
samples = tier1["samples"]
print(f"\nTier 1: {len(samples)} samples")

# ============================================================
# Q1: Class imbalance per model
# ============================================================
print("\n" + "=" * 70)
print("  Q1: CLASS IMBALANCE PER MODEL")
print("=" * 70)
models = ["llama2", "mistral", "llama31", "qwen25", "gemma2"]
for m in models:
    key = f"is_drifted_{m}"
    d = sum(1 for s in samples if s.get(key, False))
    s_ = len(samples) - d
    ratio = d / s_ if s_ > 0 else 0
    print(f"  {m:10s}: {d:5d} drifted / {s_:5d} stable  (ratio 1:{s_/d:.1f})" if d > 0 else f"  {m:10s}: 0 drifted")

# ============================================================
# Q2: What categories map to "drifted" for each model?
# ============================================================
print("\n" + "=" * 70)
print("  Q2: CATEGORY → DRIFT LABEL MAPPING (for Qwen2.5)")
print("=" * 70)
cat_drift = defaultdict(lambda: {"drifted": 0, "stable": 0})
for s in samples:
    cat = s.get("category", "unknown")
    if s.get("is_drifted_qwen25", False):
        cat_drift[cat]["drifted"] += 1
    else:
        cat_drift[cat]["stable"] += 1

print(f"  {'Category':20s} {'Drifted':>8s} {'Stable':>8s} {'%Drifted':>10s}")
print("  " + "-" * 50)
for cat in sorted(cat_drift.keys()):
    d = cat_drift[cat]["drifted"]
    s_ = cat_drift[cat]["stable"]
    pct = d / (d + s_) * 100 if (d + s_) > 0 else 0
    print(f"  {cat:20s} {d:8d} {s_:8d} {pct:9.1f}%")

print("\n  KEY QUESTION: Is known_drift labeled as drifted?")
kd_drifted = sum(1 for s in samples if s.get("category") == "known_drift" and s.get("is_drifted_qwen25", False))
kd_total = sum(1 for s in samples if s.get("category") == "known_drift")
print(f"  known_drift samples labeled drifted for Qwen: {kd_drifted}/{kd_total}")

# ============================================================
# Q3: YEAR LEAKAGE — distribution of query years by drift label
# ============================================================
print("\n" + "=" * 70)
print("  Q3: YEAR LEAKAGE CHECK (Qwen2.5)")
print("=" * 70)

# Extract year from query or from 'year' field
year_by_drift = defaultdict(lambda: {"drifted": 0, "stable": 0})
for s in samples:
    yr = s.get("year", "unknown")
    if s.get("is_drifted_qwen25", False):
        year_by_drift[yr]["drifted"] += 1
    else:
        year_by_drift[yr]["stable"] += 1

print(f"  {'Year':>6s} {'Drifted':>8s} {'Stable':>8s} {'%Drifted':>10s}")
print("  " + "-" * 36)
for yr in sorted(year_by_drift.keys(), key=lambda x: str(x)):
    d = year_by_drift[yr]["drifted"]
    s_ = year_by_drift[yr]["stable"]
    pct = d / (d + s_) * 100 if (d + s_) > 0 else 0
    print(f"  {str(yr):>6s} {d:8d} {s_:8d} {pct:9.1f}%")

# Check if "In 20XX" appears in queries
year_in_query = Counter()
for s in samples:
    q = s.get("query", "")
    for y in range(2010, 2027):
        if str(y) in q:
            year_in_query[y] += 1
            break
    else:
        year_in_query["no_year"] += 1

print(f"\n  Year mentioned in query text:")
for yr, n in sorted(year_in_query.items(), key=lambda x: str(x)):
    print(f"    {yr}: {n}")

# CRITICAL: For drifted vs stable, what years appear in the query?
print(f"\n  Query year distribution for DRIFTED vs STABLE (Qwen):")
drifted_years = Counter()
stable_years = Counter()
for s in samples:
    q = s.get("query", "")
    yr_found = None
    for y in range(2010, 2027):
        if str(y) in q:
            yr_found = y
            break
    if yr_found is None:
        yr_found = "no_year"
    if s.get("is_drifted_qwen25", False):
        drifted_years[yr_found] += 1
    else:
        stable_years[yr_found] += 1

all_years = sorted(set(list(drifted_years.keys()) + list(stable_years.keys())), key=lambda x: str(x))
print(f"  {'Year':>8s} {'Drifted':>8s} {'Stable':>8s}")
print("  " + "-" * 28)
for yr in all_years:
    print(f"  {str(yr):>8s} {drifted_years.get(yr, 0):8d} {stable_years.get(yr, 0):8d}")

# ============================================================
# Q4: What does temporal_zone filter actually include?
# ============================================================
print("\n" + "=" * 70)
print("  Q4: TEMPORAL ZONE DISTRIBUTION")
print("=" * 70)
tz_counts = Counter(s.get("temporal_zone", "none") for s in samples)
for tz, n in tz_counts.most_common():
    print(f"  {str(tz):20s}: {n:6d}")

# ============================================================
# Q5: expected_answer and model_likely_answer
# ============================================================
print("\n" + "=" * 70)
print("  Q5: MODEL_LIKELY_ANSWER FIELD")
print("=" * 70)
has_mla = sum(1 for s in samples if s.get("model_likely_answer") and str(s.get("model_likely_answer")).strip())
print(f"  Samples with model_likely_answer: {has_mla}/{len(samples)}")
if has_mla > 0:
    # Show a few examples
    count = 0
    for s in samples:
        mla = s.get("model_likely_answer", "")
        if mla and str(mla).strip():
            ea = s.get("expected_answer", "")
            q = s.get("query", "")[:60]
            print(f"    Query: {q}")
            print(f"    Expected: {ea}")
            print(f"    Model likely: {mla}")
            print(f"    Drifted (qwen): {s.get('is_drifted_qwen25', False)}")
            print()
            count += 1
            if count >= 3:
                break

# ============================================================
# Q6: Noise samples (Arabic, empty relation)
# ============================================================
print("\n" + "=" * 70)
print("  Q6: NOISE SAMPLES IN TIER 1")
print("=" * 70)
empty_rel = sum(1 for s in samples if not s.get("relation", "").strip())
arabic = sum(1 for s in samples if any(ord(c) > 0x0600 and ord(c) < 0x06FF for c in s.get("relation", "")))
tiny_rels = [(r, n) for r, n in Counter(s.get("relation", "") for s in samples).items() if n < 20]
print(f"  Empty relation: {empty_rel}")
print(f"  Arabic relation: {arabic}")
print(f"  Relations with <20 samples: {tiny_rels}")

# ============================================================
# Q7: Differential facts distribution across relations
# ============================================================
print("\n" + "=" * 70)
print("  Q7: DIFFERENTIAL FACTS BY RELATION")
print("=" * 70)
diff_by_rel = Counter()
total_by_rel = Counter()
for s in samples:
    rel = s.get("relation", "unknown")
    total_by_rel[rel] += 1
    labels = set()
    for m in models:
        labels.add(s.get(f"is_drifted_{m}", False))
    if len(labels) > 1:
        diff_by_rel[rel] += 1

n_diff = sum(diff_by_rel.values())
print(f"  Total differential facts: {n_diff}")
print(f"\n  {'Relation':30s} {'Differential':>12s} {'Total':>8s} {'%Diff':>8s}")
print("  " + "-" * 60)
for rel in sorted(total_by_rel.keys()):
    d = diff_by_rel.get(rel, 0)
    t = total_by_rel[rel]
    pct = d / t * 100 if t > 0 else 0
    print(f"  {rel:30s} {d:12d} {t:8d} {pct:7.1f}%")

# ============================================================
# Q8: Sample query format examples per category
# ============================================================
print("\n" + "=" * 70)
print("  Q8: SAMPLE QUERIES PER CATEGORY")
print("=" * 70)
for cat in ["stable", "no_drift", "known_drift", "unknown_drift"]:
    cat_samples = [s for s in samples if s.get("category") == cat]
    print(f"\n  [{cat}] ({len(cat_samples)} samples)")
    for s in cat_samples[:3]:
        d_labels = " | ".join(f"{m}={'D' if s.get(f'is_drifted_{m}', False) else 'S'}" for m in models)
        print(f"    Q: {s.get('query', '')[:80]}")
        print(f"    A: {s.get('expected_answer', '')[:40]}")
        print(f"    Year: {s.get('year', '?')}, Drift date: {str(s.get('drift_date', ''))[:10]}")
        print(f"    Labels: {d_labels}")
        print()

# ============================================================
# SUMMARY: Is year leakage a real problem?
# ============================================================
print("\n" + "=" * 70)
print("  VERDICT: YEAR LEAKAGE RISK")
print("=" * 70)
# Check if drifted samples are concentrated in recent years
drifted_recent = sum(1 for s in samples if s.get("is_drifted_qwen25", False) and int(s.get("year", 0)) >= 2024)
drifted_total = sum(1 for s in samples if s.get("is_drifted_qwen25", False))
stable_recent = sum(1 for s in samples if not s.get("is_drifted_qwen25", False) and int(s.get("year", 0)) >= 2024)
stable_total = sum(1 for s in samples if not s.get("is_drifted_qwen25", False))

print(f"  Drifted in 2024+: {drifted_recent}/{drifted_total} ({drifted_recent/drifted_total*100:.1f}%)" if drifted_total > 0 else "  No drifted samples")
print(f"  Stable in 2024+:  {stable_recent}/{stable_total} ({stable_recent/stable_total*100:.1f}%)" if stable_total > 0 else "  No stable samples")

if drifted_total > 0 and stable_total > 0:
    d_pct = drifted_recent / drifted_total
    s_pct = stable_recent / stable_total
    if d_pct > 0.8 and s_pct < 0.3:
        print("\n  ⚠️  HIGH RISK: Drifted samples are concentrated in recent years.")
        print("     The probe may be learning YEAR, not DRIFT.")
    elif d_pct > s_pct + 0.2:
        print("\n  ⚠️  MODERATE RISK: Some year-drift correlation exists.")
        print("     Paraphrase test + year-controlled subset needed.")
    else:
        print("\n  ✅ LOW RISK: Year distribution is similar across drifted/stable.")

print(f"\n{'=' * 70}")
print("  Run: python data_diagnostic.py")
print(f"{'=' * 70}")