| |
| """Gesamtbericht v3 (plain) mit VORAB festgelegten Erfolgskriterien. Rein lesend, laeuft auf der CPU.""" |
| import json |
| from pathlib import Path |
|
|
| P = Path("/home/david/Schreibtisch/Platte/models/graf_zahl") |
|
|
|
|
| def load(p): |
| try: |
| return json.load(open(p, encoding="utf-8")) |
| except Exception: |
| return None |
|
|
|
|
| def rate(rows, key="richtig"): |
| return sum(bool(r[key]) for r in rows), len(rows) |
|
|
|
|
| def main(): |
| checks = [] |
| ab = load(P / "ab_test_results_v3.json") |
| if ab: |
| a = [r for r in ab if r["adapter"] == "A (hardcore)"]; b = [r for r in ab if r["adapter"] == "B (v3)"] |
| print("=== 1. A/B 50 Woerter (Regression) ===") |
| for cat in dict.fromkeys(r["kategorie"] for r in b): |
| print(f" {cat:12} Hardcore {rate([r for r in a if r['kategorie'] == cat])} v3 {rate([r for r in b if r['kategorie'] == cat])}") |
| ok, n = rate(b) |
| print(f" GESAMT Hardcore {rate(a)} v3 {(ok, n)}") |
| checks.append(("50-Wort-Test: v3 mindestens 48/50 (keine Regression)", ok >= 48 and n == 50, f"{ok}/{n}")) |
| else: |
| checks.append(("50-Wort-Test", False, "keine Ergebnisse")) |
|
|
| gen = load(P / "generalization_results_v3.json"); old = load(P / "generalization_results.json") |
| if gen and old: |
| print("\n=== 2. Generalisierungstest (28 handgeschriebene Aufgaben), streng / Think-Zahl ===") |
| for cat in dict.fromkeys(r["kategorie"] for r in gen): |
| g = [r for r in gen if r["kategorie"] == cat]; o = [r for r in old if r["kategorie"] == cat] |
| print(f" {cat:24} Hardcore {rate(o)[0]}/{len(o)} (Think {sum(r['sicherheitszahl'] == r['laenge'] for r in o)}) " |
| f"v3 {rate(g)[0]}/{len(g)} (Think {sum(r['sicherheitszahl'] == r['laenge'] for r in g)}) abgebrochen v3: {sum(r['abgebrochen'] for r in g)}") |
| ok, n = rate(gen) |
| print(f" GESAMT streng: Hardcore {rate(old)} v3 {(ok, n)}") |
| ov = [r for r in gen if r["kategorie"].startswith("B")]; co = [r for r in gen if r["kategorie"].startswith("C")] |
| checks.append(("Generalisierung gesamt: v3 mindestens 22/28 (Hardcore 15/28)", ok >= 22, f"{ok}/{n}")) |
| checks.append(("Ueberlaenge 48-80 Zeichen: mindestens 5/7 streng (Hardcore 1/7)", rate(ov)[0] >= 5, f"{rate(ov)[0]}/{len(ov)}")) |
| checks.append(("Code mit Leerzeichen: mindestens 4/6 (Hardcore 0/6)", rate(co)[0] >= 4, f"{rate(co)[0]}/{len(co)}")) |
| else: |
| checks.append(("Generalisierungstest", False, "keine Ergebnisse")) |
|
|
| held = load(P / "v3" / "heldout_results_v3.json") |
| if held: |
| print("\n=== 3. Frisches v3-Testset (42 Aufgaben), streng / Think-Zahl ===") |
| for cat in dict.fromkeys(r["kategorie"] for r in held): |
| h = [r for r in held if r["kategorie"] == cat and r["adapter"] == "A (hardcore)"] |
| v = [r for r in held if r["kategorie"] == cat and r["adapter"] == "B (v3)"] |
| print(f" {cat:26} Hardcore {rate(h)[0]}/{len(h)} ({rate(h, 'think_richtig')[0]}) v3 {rate(v)[0]}/{len(v)} ({rate(v, 'think_richtig')[0]}) abgebr. v3: {sum(r['abgebrochen'] for r in v)}") |
| v = [r for r in held if r["adapter"] == "B (v3)"]; h = [r for r in held if r["adapter"] == "A (hardcore)"] |
| ok, n = rate(v) |
| print(f" GESAMT streng: Hardcore {rate(h)} v3 {(ok, n)}") |
| checks.append(("Frisches Testset: v3 mindestens 75 % streng", n > 0 and ok / n >= 0.75, f"{ok}/{n} = {ok / max(n, 1) * 100:.0f} %")) |
| else: |
| checks.append(("Frisches Testset", False, "keine Ergebnisse")) |
|
|
| intel = load(P / "compare_intelligence_v3_results.json") |
| if intel and "v3 plain" in intel and "Hardcore (v2-Daten)" in intel: |
| print("\n=== 4. Intelligenz und Mathe (je 100, Arithmetik 40) ===") |
| names = list(intel) |
| print(f" {'Test':15} " + " ".join(f"{n[:22]:>24}" for n in names)) |
| for key in ("GSM8K", "MMLU", "ARC-Challenge", "Arithmetik"): |
| print(f" {key:15} " + " ".join(f"{intel[n][key]['richtig']:>24}" for n in names)) |
| v, h = intel["v3 plain"], intel["Hardcore (v2-Daten)"] |
| for key, tol in (("GSM8K", 3), ("MMLU", 5), ("ARC-Challenge", 5), ("Arithmetik", 3)): |
| checks.append((f"{key}: v3 hoechstens {tol} Punkte unter Hardcore", v[key]["richtig"] >= h[key]["richtig"] - tol, |
| f"{v[key]['richtig']} gegen {h[key]['richtig']}")) |
| else: |
| checks.append(("Intelligenz/Mathe", False, "keine Ergebnisse")) |
|
|
| print("\n" + "=" * 78 + "\nERFOLGSKRITERIEN (vorab festgelegt)\n" + "=" * 78) |
| for name, ok, val in checks: |
| print(f" [{'BESTANDEN' if ok else 'DURCHGEFALLEN':13}] {name} -> {val}") |
| print(f"\n {sum(ok for _, ok, _ in checks)} von {len(checks)} Kriterien bestanden.") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
