| |
| """Frisches v3-Testset (42 Aufgaben, nie im Training): Hardcore-Adapter gegen v3-Adapter, bf16, Greedy, GPU 0.""" |
| import os |
| os.environ["CUDA_VISIBLE_DEVICES"] = "0" |
|
|
| PLATTE = "/home/david/Schreibtisch/Platte" |
| os.environ.setdefault("HF_HOME", f"{PLATTE}/cache/huggingface") |
| os.environ.setdefault("HF_HUB_CACHE", f"{PLATTE}/cache/huggingface/hub") |
|
|
| import json |
| import re |
| import time |
|
|
| import torch |
| from peft import PeftModel |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
| BASE = f"{PLATTE}/models/graf_zahl/qwen_healed_merged" |
| |
| ADAPTERS = { |
| "A (hardcore)": (f"{PLATTE}/models/graf_zahl/qwen_graf_zahl_hardcore_lora", 2500), |
| "B (v3)": (f"{PLATTE}/models/graf_zahl/qwen_graf_zahl_v3_plain_lora", 4500), |
| } |
| HELD = f"{PLATTE}/models/graf_zahl/v3/heldout_v3.json" |
| RESULTS = f"{PLATTE}/models/graf_zahl/v3/heldout_results_v3.json" |
| DATASET = "/mnt/models/graf_zahl/dataset_massive.jsonl" |
| SPACE = "␣" |
|
|
|
|
| def analyse(s, answer): |
| pointers = re.findall(r"\[Position (\d+): (.)\]", answer) |
| chars = "".join((" " if c == SPACE else c) for _, c in pointers) |
| final = re.findall(r"Das sind (\d+)", answer) |
| safety = re.findall(r"Zur Sicherheit nachgezählt:\s*(\d+)", answer) |
| return {"endzahl": int(final[-1]) if final else None, "sicherheitszahl": int(safety[-1]) if safety else None, |
| "pointer_exakt": chars == s, "letzte_position": int(pointers[-1][0]) if pointers else 0} |
|
|
|
|
| def main(): |
| held = json.load(open(HELD, encoding="utf-8")) |
| system = json.loads(open(DATASET, encoding="utf-8").readline())["messages"][0]["content"] |
| tok = AutoTokenizer.from_pretrained(BASE) |
| model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16, device_map={"": 0}) |
| first = True |
| for name, (path, _) in ADAPTERS.items(): |
| if first: |
| model = PeftModel.from_pretrained(model, path, adapter_name=name[0]); first = False |
| else: |
| model.load_adapter(path, adapter_name=name[0]) |
| model.eval() |
| print(f"{len(held)} Aufgaben, Adapter {list(ADAPTERS)}", flush=True) |
|
|
| results, t0 = [], time.time() |
| for name, (_, max_new) in ADAPTERS.items(): |
| model.set_adapter(name[0]) |
| for h in held: |
| s = h["text"] |
| messages = [{"role": "system", "content": system}, {"role": "user", "content": h["prompt"]}] |
| text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| inputs = tok(text, return_tensors="pt", add_special_tokens=False).to("cuda") |
| with torch.no_grad(): |
| out = model.generate(**inputs, max_new_tokens=max_new, do_sample=False, temperature=0.0, top_p=1.0, |
| eos_token_id=tok.eos_token_id, pad_token_id=tok.pad_token_id) |
| new = out[0][inputs["input_ids"].shape[1]:] |
| answer = tok.decode(new, skip_special_tokens=True) |
| a = analyse(s, answer) |
| rec = {"adapter": name, "kategorie": h["kategorie"], "text": s, "laenge": len(s), "prompt": h["prompt"], |
| "tokens": len(new), "abgebrochen": int(new[-1]) != tok.eos_token_id, "richtig": a["endzahl"] == len(s), |
| "think_richtig": a["sicherheitszahl"] == len(s), "antwort": answer, **a} |
| results.append(rec) |
| print(f"[{name}] {h['kategorie'][:22]:22} len {len(s):2} -> {str(a['endzahl']):>4} " |
| f"{'ABGEBR.' if rec['abgebrochen'] else ('ok' if rec['richtig'] else 'FALSCH'):8} ({len(new)} Tok, {time.time() - t0:.0f}s)", flush=True) |
| json.dump(results, open(RESULTS, "w", encoding="utf-8"), ensure_ascii=False) |
|
|
| print("\n===== AUSWERTUNG (streng 'Das sind N' / Think-Zahl) =====") |
| cats = list(dict.fromkeys(h["kategorie"] for h in held)) |
| print(f"{'Kategorie':26} {'n':>3} " + " ".join(f"{a:>20}" for a in ADAPTERS)) |
| for cat in cats + ["Gesamt"]: |
| row = [] |
| for a in ADAPTERS: |
| rs = [r for r in results if r["adapter"] == a and (cat == "Gesamt" or r["kategorie"] == cat)] |
| row.append(f"{sum(r['richtig'] for r in rs):>3}/{len(rs):<3} / {sum(r['think_richtig'] for r in rs):>3}/{len(rs):<3}") |
| n = len([h for h in held if cat == "Gesamt" or h["kategorie"] == cat]) |
| print(f"{cat:26} {n:>3} " + " ".join(f"{c:>20}" for c in row)) |
| print(f"\nErgebnisse: {RESULTS}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
