Blame for code/compare_intelligence_v3.py

#!/usr/bin/env python3
"""Intelligenz- und Mathetests: Original Qwen2.5-7B-Instruct gegen geheiltes Modell gegen Hardcore-Modell.

Gleiche Aufgaben, gleiche Einstellungen (bf16, GPU 0, Greedy, repetition_penalty 1.0, Standard-Chat-Template):
  GSM8K (100, Rechenaufgaben, Antwort per Text) | MMLU (100, Multiple Choice, Logits ueber A-D)
  ARC-Challenge (100, Multiple Choice) | Arithmetik (40, selbst erzeugt).
"""
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 gc
import json
import random
import re
import time

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

MODELS = [
    ("Original Qwen2.5-7B-Instruct", "unsloth/Qwen2.5-7B-Instruct"),
    ("Geheilt (ohne LoRA)", f"{PLATTE}/models/graf_zahl/qwen_healed_merged"),
    ("Hardcore (v2-Daten)", f"{PLATTE}/models/graf_zahl/qwen_graf_zahl_hardcore_merged"),
    ("v3 plain", f"{PLATTE}/models/graf_zahl/qwen_graf_zahl_v3_plain_merged"),
]
RESULTS = f"{PLATTE}/models/graf_zahl/compare_intelligence_v3_results.json"
N_GSM, N_MMLU, N_ARC = 100, 100, 100
SEED = 3407
BATCH = 16


def build_tasks():
    rng = random.Random(SEED)
    g = load_dataset("openai/gsm8k", "main", split="test")
    gsm = [{"q": g[i]["question"], "a": float(g[i]["answer"].split("####")[-1].strip().replace(",", ""))}
           for i in rng.sample(range(len(g)), N_GSM)]
    m = load_dataset("cais/mmlu", "all", split="test")
    mmlu = []
    for i in rng.sample(range(len(m)), N_MMLU):
        r = m[i]
        opts = "\n".join(f"{l}. {c}" for l, c in zip("ABCD", r["choices"]))
        mmlu.append({"q": f"The following is a multiple choice question about {r['subject'].replace('_', ' ')}.\n\n"
                          f"{r['question']}\n{opts}\n\nAnswer with only the letter of the correct option.",
                     "a": "ABCD"[r["answer"]], "labels": "ABCD", "subject": r["subject"]})
    a = load_dataset("allenai/ai2_arc", "ARC-Challenge", split="test")
    arc = []
    for i in rng.sample(range(len(a)), N_ARC):
        r = a[i]
        labels = r["choices"]["label"]
        opts = "\n".join(f"{l}. {t}" for l, t in zip(labels, r["choices"]["text"]))
        arc.append({"q": f"The following is a multiple choice science question.\n\n{r['question']}\n{opts}\n\n"
                         "Answer with only the letter of the correct option.", "a": r["answerKey"], "labels": "".join(labels)})
    ar = []
    for _ in range(10): x, y = rng.randint(10000, 99999), rng.randint(10000, 99999); ar.append((f"{x} + {y}", x + y, "add"))
    for _ in range(10): x, y = rng.randint(50000, 99999), rng.randint(10000, 49999); ar.append((f"{x} - {y}", x - y, "sub"))
    for _ in range(10): x, y = rng.randint(100, 999), rng.randint(11, 99); ar.append((f"{x} * {y}", x * y, "mul 3x2"))
    for _ in range(5): x, y = rng.randint(100, 999), rng.randint(100, 999); ar.append((f"{x} * {y}", x * y, "mul 3x3"))
    for _ in range(5): y, q = rng.randint(12, 99), rng.randint(100, 999); ar.append((f"{y * q} / {y}", q, "div"))
    arith = [{"q": f"Berechne: {e} = ?\nAntworte nur mit der Zahl.", "a": float(v), "typ": t} for e, v, t in ar]
    return {"GSM8K": gsm, "MMLU": mmlu, "ARC-Challenge": arc, "Arithmetik": arith}


def to_prompt(tok, text):
    return tok.apply_chat_template([{"role": "user", "content": text}], tokenize=False, add_generation_prompt=True)


def last_number(s):
    nums = re.findall(r"-?\d[\d,]*\.?\d*", s.replace("$", ""))
    return float(nums[-1].replace(",", "").rstrip(".")) if nums else None


def gsm_number(s):
    m = re.findall(r"Answer:\s*\$?\s*(-?\d[\d,]*\.?\d*)", s)
    return float(m[-1].replace(",", "").rstrip(".")) if m else last_number(s)


@torch.no_grad()
def generate_batch(model, tok, prompts, max_new):
    outs = []
    order = sorted(range(len(prompts)), key=lambda i: len(prompts[i]))
    res = [None] * len(prompts)
    for s in range(0, len(order), BATCH):
        idx = order[s:s + BATCH]
        enc = tok([prompts[i] for i in idx], return_tensors="pt", padding=True, add_special_tokens=False).to("cuda")
        out = model.generate(**enc, max_new_tokens=max_new, do_sample=False, repetition_penalty=1.0,
                             eos_token_id=tok.eos_token_id, pad_token_id=tok.pad_token_id)
        for j, i in enumerate(idx):
            res[i] = tok.decode(out[j][enc["input_ids"].shape[1]:], skip_special_tokens=True)
    return res


@torch.no_grad()
def choose_letters(model, tok, prompts, labels_list):
    letter_ids = {}
    for l in set("".join(labels_list)):
        ids = tok.encode(l, add_special_tokens=False)
        assert len(ids) == 1, (l, ids)
        letter_ids[l] = ids[0]
    res = [None] * len(prompts)
    order = sorted(range(len(prompts)), key=lambda i: len(prompts[i]))
    for s in range(0, len(order), BATCH):
        idx = order[s:s + BATCH]
        enc = tok([prompts[i] for i in idx], return_tensors="pt", padding=True, add_special_tokens=False).to("cuda")
        logits = model(**enc).logits[:, -1, :].float()
        for j, i in enumerate(idx):
            res[i] = max(labels_list[i], key=lambda l: logits[j, letter_ids[l]].item())
    return res


def main():
    tasks = build_tasks()
    print({k: len(v) for k, v in tasks.items()}, flush=True)
    results = {}
    for name, path in MODELS:
        t0 = time.time()
        tok = AutoTokenizer.from_pretrained(path)
        tok.padding_side = "left"
        model = AutoModelForCausalLM.from_pretrained(path, dtype=torch.bfloat16, device_map={"": 0}).eval()
        print(f"\n===== {name} geladen ({len(tok)} Tokens im Vokabular, {time.time() - t0:.0f}s) =====", flush=True)
        r = {}

        outs = generate_batch(model, tok, [to_prompt(tok, t["q"] + "\nSolve the problem step by step. End your reply with 'Answer: <number>'.")
                                           for t in tasks["GSM8K"]], 512)
        ok = [gsm_number(o) is not None and abs(gsm_number(o) - t["a"]) < 1e-6 for o, t in zip(outs, tasks["GSM8K"])]
        r["GSM8K"] = {"richtig": sum(ok), "n": len(ok), "beispiel_antwort": outs[0][-300:]}
        print(f"  GSM8K: {sum(ok)}/{len(ok)}  ({time.time() - t0:.0f}s)", flush=True)

        for key in ("MMLU", "ARC-Challenge"):
            pick = choose_letters(model, tok, [to_prompt(tok, t["q"]) for t in tasks[key]], [t["labels"] for t in tasks[key]])
            ok = [p == t["a"] for p, t in zip(pick, tasks[key])]
            r[key] = {"richtig": sum(ok), "n": len(ok)}
            print(f"  {key}: {sum(ok)}/{len(ok)}  ({time.time() - t0:.0f}s)", flush=True)

        outs = generate_batch(model, tok, [to_prompt(tok, t["q"]) for t in tasks["Arithmetik"]], 40)
        ok = [last_number(o) is not None and abs(last_number(o) - t["a"]) < 1e-6 for o, t in zip(outs, tasks["Arithmetik"])]
        by = {}
        for t, k in zip(tasks["Arithmetik"], ok):
            by.setdefault(t["typ"], []).append(k)
        r["Arithmetik"] = {"richtig": sum(ok), "n": len(ok), "je_typ": {k: f"{sum(v)}/{len(v)}" for k, v in by.items()}}
        print(f"  Arithmetik: {sum(ok)}/{len(ok)} {r['Arithmetik']['je_typ']}  ({time.time() - t0:.0f}s)", flush=True)

        results[name] = r
        json.dump(results, open(RESULTS, "w", encoding="utf-8"), ensure_ascii=False, indent=1)
        del model
        gc.collect()
        torch.cuda.empty_cache()

    print("\n===== AUSWERTUNG =====")
    names = list(results)
    print(f"{'Test':16} " + " ".join(f"{n[:28]:>30}" for n in names))
    for key in ("GSM8K", "MMLU", "ARC-Challenge", "Arithmetik"):
        print(f"{key:16} " + " ".join(f"{results[n][key]['richtig']:>4}/{results[n][key]['n']:<3} = {results[n][key]['richtig'] / results[n][key]['n'] * 100:5.1f} %".rjust(30) for n in names))
    print(f"\nErgebnisse: {RESULTS}")


if __name__ == "__main__":
    main()