| |
| """Schwerer Zaehl-Vergleich (Generalisierungs-Set: Sonderzeichen, Ueberlaenge, Code, andere Prompts). |
| |
| Dieselben 28 Aufgaben und Prompts wie beim Generalisierungstest unseres Modells, gestellt an QuiteQue-Modelle. |
| """ |
| import json |
| import sys |
| import threading |
| import time |
| from concurrent.futures import ThreadPoolExecutor |
|
|
| sys.path.insert(0, "/mnt/models/graf_zahl") |
| from compare_counting_quiteque import ALLOWED, ask, last_int |
|
|
| PLATTE = "/home/david/Schreibtisch/Platte" |
| GEN = f"{PLATTE}/models/graf_zahl/generalization_results.json" |
| OUT = f"{PLATTE}/models/graf_zahl/compare_counting_hard_results.json" |
| SUFFIX = " Antworte nur mit der Zahl." |
|
|
| MODELS = [ |
| ("50", "T79 Qwen3 4B Instruct (CPU)", 1), |
| ("gemma-12b-chat", "Gemma 4 12B (lokal)", 2), |
| ("120", "PropellerA Qwen3.8 27B (lokal)", 1), |
| ("or-liquid/lfm-2.5-2.6b:free", "Liquid LFM 2.5 2.6B (free)", 2), |
| ("or-google/gemma-4-31b-it:free", "Gemma 4 31B (free)", 2), |
| ("or-nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free", "Nemotron 3 Nano 30B Reasoning (free)", 2), |
| ("or-nex-agi/nex-n2.5-pro:free", "Nex-N2.5 Pro (free)", 2), |
| ("or-nvidia/nemotron-3.5-lightning:free", "Nemotron 3.5 Lightning (free)", 2), |
| ] |
|
|
|
|
| def main(): |
| assert all(ALLOWED(m[0]) for m in MODELS) |
| ours = json.load(open(GEN, encoding="utf-8")) |
| items = [(r["kategorie"], r["text"], len(r["text"]), r["prompt"] + SUFFIX, r["endzahl"]) for r in ours] |
| cats = list(dict.fromkeys(i[0] for i in items)) |
| print(f"{len(items)} Aufgaben, {len(MODELS)} Modelle", flush=True) |
| results, lock = {}, threading.Lock() |
|
|
| for mid, name, threads in MODELS: |
| rows = [] |
| def one(it): |
| cat, s, n, prompt, _ = it |
| content, err = ask(mid, prompt, None, max_tokens=6000) |
| got = last_int(content) if not err else None |
| return {"kategorie": cat, "text": s, "soll": n, "antwort": content[:300], "zahl": got, "richtig": got == n, "fehler": err} |
| t0 = time.time() |
| with ThreadPoolExecutor(threads) as ex: |
| for r in ex.map(one, items): |
| rows.append(r) |
| with lock: |
| results[mid] = {"name": name, "rows": rows} |
| json.dump(results, open(OUT, "w", encoding="utf-8"), ensure_ascii=False) |
| print(f"[{name}] fertig: {sum(r['richtig'] for r in rows)}/{len(rows)} richtig, {sum(1 for r in rows if r['fehler'])} Fehler, {time.time() - t0:.0f}s", flush=True) |
|
|
| ours_rows = [{"kategorie": c, "richtig": e == n, "soll": n} for c, s, n, p, e in items] |
| results["unser"] = {"name": "UNSER Hardcore-Modell ('Das sind N')", "rows": ours_rows} |
| |
| think = {"kategorie": None} |
| ours_think = [{"kategorie": r["kategorie"], "richtig": r["sicherheitszahl"] == len(r["text"])} for r in ours] |
| results["unser_think"] = {"name": "UNSER Modell (Zahl im Think-Block)", "rows": ours_think} |
| json.dump(results, open(OUT, "w", encoding="utf-8"), ensure_ascii=False) |
|
|
| print("\n===== AUSWERTUNG (richtig je Kategorie) =====") |
| print(f"{'Modell':44} " + " ".join(f"{c[:10]:>11}" for c in cats) + f" {'Gesamt':>8}") |
| for key in ["unser", "unser_think"] + [m[0] for m in MODELS]: |
| rows = results[key]["rows"] |
| per = [(sum(r["richtig"] for r in rows if r["kategorie"] == c), len([r for r in rows if r["kategorie"] == c])) for c in cats] |
| print(f"{results[key]['name'][:44]:44} " + " ".join(f"{a:>7}/{b:<3}" for a, b in per) + f" {sum(a for a, _ in per):>4}/{len(rows)}") |
| print(f"\nErgebnisse: {OUT}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
