code/Abforge_Training/scripts/v17_case_check.py

#!/usr/bin/env python3
"""Show a few concrete papers — v15 actual n_pred, v16 actual n_pred, n_gt,
and what each setting would charge.

Stratified pick: small-GT, mid-GT, large-GT to show edge behavior.
"""
import json
import re
from pathlib import Path

ROOT = Path("/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training")

_TARGET_BULLET_RE = re.compile(
    r'\n\s*[-*]\s*(?:\*\*)?Target Module(?:\*\*)?\s*:', re.IGNORECASE)
_RQ_BULLET_RE = re.compile(
    r'\n\s*[-*]\s*(?:\*\*)?Research Question(?:\*\*)?\s*:', re.IGNORECASE)


def extract_result(raw):
    m = re.search(r"<Result>(.*?)</Result>", raw, re.DOTALL | re.IGNORECASE)
    if m: return m.group(1)
    m2 = re.search(r"<Result>(.*)", raw, re.DOTALL | re.IGNORECASE)
    return m2.group(1) if m2 else raw


def count_pairs(raw):
    t = "\n" + extract_result(raw or "")
    return min(len(_TARGET_BULLET_RE.findall(t)), len(_RQ_BULLET_RE.findall(t)))


def cp_progressive(n, g, tol, base, step, cap=0.5):
    excess = max(0, n - g - tol)
    if excess <= 0: return 0.0
    return min(cap, sum(base + step * (i - 1) for i in range(1, excess + 1)))


def cp_fixed(n, g, tol, rate, cap=0.5):
    return min(cap, max(0, n - g - tol) * rate)


SETTINGS = [
    ("v16-now", lambda n, g: cp_progressive(n, g, 1, 0.05, 0.01)),
    ("E (rec)", lambda n, g: cp_progressive(n, g, 2, 0.01, 0.01)),
    ("A      ", lambda n, g: cp_progressive(n, g, 1, 0.01, 0.01)),
    ("C fixed", lambda n, g: cp_fixed(n, g, 1, 0.02)),
]


def title_key(r):
    return r.get("meta", {}).get("title", "")


v15 = {title_key(json.loads(l)): json.loads(l)
       for l in open(ROOT / "infer/task1_v15_ckpt100_bench50_qwen3_infer_task1.jsonl")}
v16 = {title_key(json.loads(l)): json.loads(l)
       for l in open(ROOT / "infer/task1_v16_ckpt50_bench50_qwen3_infer_task1.jsonl")}
evals = [json.loads(l) for l in open(ROOT / "infer/task1_v15_ckpt100_eval_50.jsonl")]

papers = []
for ev in evals:
    title = title_key(ev)
    if title not in v15 or title not in v16: continue
    n_gt = ev.get("n_gt", 0)
    n15 = count_pairs(v15[title].get("infer_task1_response", ""))
    n16 = count_pairs(v16[title].get("infer_task1_response", ""))
    mr = ev.get("match_rate", 0)
    papers.append((title, n_gt, n15, n16, mr))

# stratified pick: 2 small-GT, 2 mid-GT, 2 large-GT
papers.sort(key=lambda x: x[1])
small = papers[:2] + papers[3:5]
mid = [p for p in papers if 3 <= p[1] <= 4][:3]
large = [p for p in papers if p[1] >= 6][:3]

print("CASE-BY-CASE count_penalty under each setting")
print("=" * 100)
print(f"{'title':<55} {'n_gt':>4} {'n15':>4} {'n16':>4}   "
      f"{'v16-now (v15→v16)':>22}  {'E (v15→v16)':>16}  {'A (v15→v16)':>16}  {'C (v15→v16)':>16}")
print("-" * 160)

for label, group in [("SMALL n_gt", small), ("MID n_gt", mid), ("LARGE n_gt", large)]:
    print(f"\n--- {label} ---")
    for title, n_gt, n15, n16, mr in group:
        title_s = (title[:50] + "…") if len(title) > 50 else title
        row = f"{title_s:<55} {n_gt:>4} {n15:>4} {n16:>4}  "
        for name, fn in SETTINGS:
            p15 = fn(n15, n_gt)
            p16 = fn(n16, n_gt)
            row += f"  {p15:.3f}→{p16:.3f}    "
        print(row)

# also: average penalty if model "stabilizes at n=5" vs current v16 n=4
print("\n\n=== summary: avg penalty across all 49 papers if model writes constant N ===")
print(f"{'setting':<10}  {'n=4 (v16 cur)':>15}  {'n=5':>10}  {'n=6':>10}  {'n=7':>10}")
for name, fn in SETTINGS:
    row = f"{name:<10}  "
    for N in [4, 5, 6, 7]:
        mean = sum(fn(N, p[1]) for p in papers) / len(papers)
        row += f"  {mean:>10.4f}  "
    print(row)