scripts/generate_dataset_report.py

"""
generate_dataset_report.py
==========================
Automated dataset diagnostics report for GomParam-v1.
Computes difficulty calibration, lexical diversity, distractor similarity,
distribution summaries, and duplicate detection.

Outputs:
  - results/dataset_report.md  (human-readable markdown)
  - results/dataset_report.json (machine-readable)
"""

import json
import math
import re
import statistics
import unicodedata
from collections import Counter, defaultdict
from pathlib import Path

import numpy as np

DATA_DIR = Path("/mnt/data/projects/GomParam-v1/data")
OUT_DIR = Path("/mnt/data/projects/GomParam-v1/results")
OUT_DIR.mkdir(parents=True, exist_ok=True)

# ──────────────────────────────────────────────
# Load all items
# ──────────────────────────────────────────────
print("Loading dataset from local JSON files...")
all_items = []
module_counts = Counter()

for json_file in sorted(DATA_DIR.glob("*.json")):
    module_name = json_file.stem
    with open(json_file, "r", encoding="utf-8") as f:
        data = json.load(f)
    for item in data:
        item["module"] = module_name
        all_items.append(item)
    module_counts[module_name] = len(data)

N = len(all_items)
print(f"Loaded {N} items across {len(module_counts)} modules.\n")

# ──────────────────────────────────────────────
# 1. Question & Option Length Statistics
# ──────────────────────────────────────────────
print("Computing length statistics...")

q_lengths = []
correct_lengths = []
incorrect_lengths = []
all_option_lengths = []

for item in all_items:
    q = item.get("context", "") or ""
    question = item.get("question", "") or ""
    combined = f"{q} {question}".strip()
    q_lengths.append(len(combined))

    candidates = item.get("candidates", [])
    correct_idx = item.get("correct", -1)

    for i, c in enumerate(candidates):
        c_len = len(str(c))
        all_option_lengths.append(c_len)
        if i == correct_idx:
            correct_lengths.append(c_len)
        else:
            incorrect_lengths.append(c_len)


def dist_stats(values, label=""):
    """Compute mean, median, std, min, max, p95."""
    if not values:
        return {}
    arr = np.array(values)
    return {
        "mean": round(float(np.mean(arr)), 2),
        "median": round(float(np.median(arr)), 2),
        "std": round(float(np.std(arr)), 2),
        "min": int(np.min(arr)),
        "max": int(np.max(arr)),
        "p95": round(float(np.percentile(arr, 95)), 2),
    }


q_stats = dist_stats(q_lengths)
correct_stats = dist_stats(correct_lengths)
incorrect_stats = dist_stats(incorrect_lengths)

# ──────────────────────────────────────────────
# 2. Answer Distribution & Entropy
# ──────────────────────────────────────────────
print("Computing answer distribution...")

answer_dist = Counter()
for item in all_items:
    idx = item.get("correct", -1)
    label = {0: "A", 1: "B", 2: "C", 3: "D"}.get(idx, "X")
    answer_dist[label] += 1

total_ans = sum(answer_dist.values())
answer_pcts = {k: round(v / total_ans * 100, 1) for k, v in sorted(answer_dist.items())}

# Shannon entropy
probs = [v / total_ans for v in answer_dist.values() if v > 0]
entropy = -sum(p * math.log2(p) for p in probs)
max_entropy = math.log2(len(probs))

# ──────────────────────────────────────────────
# 3. Vocabulary & Lexical Diversity
# ──────────────────────────────────────────────
print("Computing lexical diversity...")

all_tokens = []
devanagari_re = re.compile(r'[\u0900-\u097F]+')

for item in all_items:
    for field in ["context", "question"]:
        text = item.get(field, "") or ""
        all_tokens.extend(devanagari_re.findall(text))
    for c in item.get("candidates", []):
        all_tokens.extend(devanagari_re.findall(str(c)))

total_tokens = len(all_tokens)
unique_tokens = len(set(all_tokens))
ttr = round(unique_tokens / total_tokens * 100, 2) if total_tokens > 0 else 0

# MTLD approximation (simplified)
def compute_mtld(tokens, threshold=0.72):
    """Measure of Textual Lexical Diversity (McCarthy & Jarvis, 2010)."""
    def mtld_forward(tokens, threshold):
        factors = 0
        types = set()
        token_count = 0
        for t in tokens:
            types.add(t)
            token_count += 1
            if len(types) / token_count < threshold:
                factors += 1
                types = set()
                token_count = 0
        if token_count > 0:
            ttr = len(types) / token_count
            factors += (1.0 - ttr) / (1.0 - threshold)
        return len(tokens) / factors if factors > 0 else len(tokens)

    fwd = mtld_forward(tokens, threshold)
    bwd = mtld_forward(tokens[::-1], threshold)
    return round((fwd + bwd) / 2, 2)

mtld = compute_mtld(all_tokens) if len(all_tokens) > 100 else 0

# ──────────────────────────────────────────────
# 4. Distractor Similarity
# ──────────────────────────────────────────────
print("Computing distractor similarity...")

def jaccard(s1, s2):
    set1 = set(str(s1).split())
    set2 = set(str(s2).split())
    if not set1 or not set2:
        return 0.0
    return len(set1 & set2) / len(set1 | set2)

pairwise_sims = []
for item in all_items:
    candidates = item.get("candidates", [])
    for i in range(len(candidates)):
        for j in range(i + 1, len(candidates)):
            sim = jaccard(candidates[i], candidates[j])
            pairwise_sims.append(sim)

sim_stats = dist_stats([s * 100 for s in pairwise_sims])  # as percentages

# ──────────────────────────────────────────────
# 5. Duplicate Detection (TF-IDF Cosine)
# ──────────────────────────────────────────────
print("Computing near-duplicate detection...")

combined_texts = []
for item in all_items:
    ctx = item.get("context", "") or ""
    q = item.get("question", "") or ""
    combined_texts.append(f"{ctx} {q}".strip())

# Simple character n-gram overlap for duplicate detection
def char_ngrams(text, n=4):
    return set(text[i:i+n] for i in range(len(text) - n + 1))

high_sim_pairs = 0
THRESHOLD = 0.90
# Sample-based: check first 500 items against each other (O(n^2) is expensive for 3000)
sample_size = min(500, N)
for i in range(sample_size):
    ng_i = char_ngrams(combined_texts[i])
    if not ng_i:
        continue
    for j in range(i + 1, sample_size):
        ng_j = char_ngrams(combined_texts[j])
        if not ng_j:
            continue
        overlap = len(ng_i & ng_j) / len(ng_i | ng_j)
        if overlap > THRESHOLD:
            high_sim_pairs += 1

# ──────────────────────────────────────────────
# 6. Category Balance
# ──────────────────────────────────────────────
print("Computing category balance...")

cat_entropy_probs = [c / N for c in module_counts.values()]
cat_entropy = -sum(p * math.log2(p) for p in cat_entropy_probs if p > 0)
max_cat_entropy = math.log2(len(module_counts))

# Difficulty distribution
difficulty_dist = Counter()
for item in all_items:
    d = item.get("difficulty", "unknown")
    difficulty_dist[d] += 1

# ──────────────────────────────────────────────
# 7. Build Report
# ──────────────────────────────────────────────
report = {
    "total_items": N,
    "total_modules": len(module_counts),
    "module_counts": dict(module_counts.most_common()),
    "question_length": q_stats,
    "correct_option_length": correct_stats,
    "incorrect_option_length": incorrect_stats,
    "answer_distribution": answer_pcts,
    "answer_entropy": round(entropy, 4),
    "max_entropy": round(max_entropy, 4),
    "vocabulary": {
        "total_tokens": total_tokens,
        "unique_types": unique_tokens,
        "ttr_pct": ttr,
        "mtld": mtld,
    },
    "distractor_similarity_pct": sim_stats,
    "near_duplicates": {
        "threshold": THRESHOLD,
        "sample_size": sample_size,
        "pairs_above_threshold": high_sim_pairs,
    },
    "category_entropy": round(cat_entropy, 4),
    "max_category_entropy": round(max_cat_entropy, 4),
    "difficulty_distribution": dict(difficulty_dist),
}

# Save JSON
with open(OUT_DIR / "dataset_report.json", "w") as f:
    json.dump(report, f, indent=2, ensure_ascii=False)

# ──────────────────────────────────────────────
# 8. Generate Markdown Report
# ──────────────────────────────────────────────
md = []
md.append("# GomParam-v1 Dataset Diagnostics Report\n")
md.append(f"**Total Items:** {N}  ")
md.append(f"**Total Modules:** {len(module_counts)}\n")

md.append("## Module Balance\n")
md.append("| Module | Count |")
md.append("|--------|-------|")
for mod, cnt in module_counts.most_common():
    md.append(f"| {mod} | {cnt} |")
md.append(f"\n**Category Entropy:** {cat_entropy:.4f} / {max_cat_entropy:.4f} (max)\n")

md.append("## Question Length\n")
md.append("| Stat | Value |")
md.append("|------|-------|")
for k, v in q_stats.items():
    md.append(f"| {k} | {v} |")

md.append("\n## Option Length (Correct vs. Incorrect)\n")
md.append("| Stat | Correct | Incorrect |")
md.append("|------|---------|-----------|")
for k in correct_stats:
    md.append(f"| {k} | {correct_stats[k]} | {incorrect_stats.get(k, 'N/A')} |")

md.append("\n## Answer Distribution\n")
md.append("| Option | Percentage |")
md.append("|--------|------------|")
for k, v in sorted(answer_pcts.items()):
    md.append(f"| {k} | {v}% |")
md.append(f"\n**Shannon Entropy:** {entropy:.4f} / {max_entropy:.4f} (max)\n")

md.append("## Vocabulary & Lexical Diversity\n")
md.append(f"- **Total Tokens:** {total_tokens:,}")
md.append(f"- **Unique Types:** {unique_tokens:,}")
md.append(f"- **Type-Token Ratio:** {ttr}%")
md.append(f"- **MTLD:** {mtld}\n")

md.append("## Distractor Similarity (Jaccard %)\n")
md.append("| Stat | Value |")
md.append("|------|-------|")
for k, v in sim_stats.items():
    md.append(f"| {k} | {v}% |")

md.append(f"\n## Near-Duplicate Detection\n")
md.append(f"- **Threshold:** {THRESHOLD}")
md.append(f"- **Sample Size:** {sample_size}")
md.append(f"- **Pairs Above Threshold:** {high_sim_pairs}\n")

md.append("## Difficulty Distribution\n")
md.append("| Difficulty | Count |")
md.append("|------------|-------|")
for d, c in difficulty_dist.most_common():
    md.append(f"| {d} | {c} |")

md.append("\n---\n*Auto-generated by `generate_dataset_report.py`*\n")

with open(OUT_DIR / "dataset_report.md", "w") as f:
    f.write("\n".join(md))

print(f"\nReport saved to:")
print(f"  {OUT_DIR / 'dataset_report.json'}")
print(f"  {OUT_DIR / 'dataset_report.md'}")
print("Done.")