Blame for code/v4_data.py

#!/usr/bin/env python3
"""v4-Zaehldaten (deterministisch): Zeichen zaehlen, Buchstabenhaeufigkeit, Buchstaben insgesamt, Woerter zaehlen.

Ohne System-Prompt, ohne Persona, ohne Zaehlshow: <think> mit Spreizung/Zaehlung, danach ein normaler Satz.
Jedes Beispiel wird gegen die bekannte Wahrheit validiert. Testtexte (alle Test-Sets) sind ausgeschlossen.

Ausgabe: v4/counting_v4.jsonl (Felder: messages, kind, text). Aufruf: v4_data.py [--selftest]
"""
import json
import os
import random
import re
import string
import sys
from collections import Counter
from pathlib import Path

os.environ["CUDA_VISIBLE_DEVICES"] = ""
sys.path.insert(0, "/mnt/models/graf_zahl")
from generate_hardcore_massive import TC, TO  # noqa: E402
from v3_data import (SPACE, code_like, compound, load_sentences, misc_string, rand_alnum, rand_sym, runs, sym)  # noqa: E402
import v4_tests  # noqa: E402

PLATTE = "/home/david/Schreibtisch/Platte"
V4 = Path(f"{PLATTE}/models/graf_zahl/v4")
SEED = 44
COUNTS = {"len_old": 1900, "len_de_word": 700, "len_de_sent": 500, "len_en_sent": 330, "len_en_word": 250, "len_code_misc": 320,
          "len_periodic": 480, "len_long": 420, "letter": 1500, "letters_all": 350, "words": 400}


# ----------------------------------------------------------------------------- Think-Bloecke
def think_len(text):
    n = len(text)
    lines = [TO, f"Wort: {text}", "Spreizung: " + " - ".join(sym(c) for c in text), "Zählung:"]
    lines += [f"[Position {i}: {sym(c)}] -> Zähler: {i}." for i, c in enumerate(text, 1)]
    lines += [f"Zur Sicherheit nachgezählt: {n}.", TC]
    return "\n".join(lines)


def think_tally(text, label, pred):
    tally, lines = 0, [TO, f"Wort: {text}", f"Gesucht: {label}", "Spreizung: " + " - ".join(sym(c) for c in text), "Zählung:"]
    for i, c in enumerate(text, 1):
        hit = pred(c)
        tally += hit
        lines.append(f"[Position {i}: {sym(c)}] -> {'ja!' if hit else 'nein.'} Bisher: {tally}.")
    lines += [f"Zur Sicherheit nachgezählt: {tally}.", TC]
    return "\n".join(lines), tally


def think_words(text):
    words = text.split()
    lines = [TO, f"Satz: {text}", "Zählung nach Wörtern (durch Leerzeichen getrennt):"]
    lines += [f"[Wort {i}: {w}] -> Zähler: {i}." for i, w in enumerate(words, 1)]
    lines += [f"Zur Sicherheit nachgezählt: {len(words)}.", TC]
    return "\n".join(lines), len(words)


# ----------------------------------------------------------------------------- Antwort-Saetze
def echo_ok(text):
    return len(text) <= 30 and "„" not in text and "“" not in text and "\n" not in text


def ans_len(text, n, lang, letters_word):
    unit_de = "Buchstaben" if letters_word else "Zeichen"
    unit_en = "letters" if letters_word else "characters"
    if lang == "de":
        if " " not in text and text.isalpha() and echo_ok(text):
            return f"Das Wort „{text}“ hat {n} {unit_de}."
        if " " in text:
            return f"Der Text hat {n} {unit_de} (Leerzeichen eingerechnet)." if not letters_word else f"Der Text hat {n} Zeichen (Leerzeichen eingerechnet)."
        return f"Der String hat {n} Zeichen."
    if " " not in text and text.isalpha() and echo_ok(text):
        return f"The word “{text}” has {n} {unit_en}."
    if " " in text:
        return f"The text has {n} characters (spaces included)."
    return f"The string has {n} characters."


def cap(s):
    return s[:1].upper() + s[1:]


def ans_letter(text, c, k, lang):
    what_de = "die Ziffer" if c.isdigit() else "der Buchstabe"
    what_en = "digit" if c.isdigit() else "letter"
    if lang == "de":
        if echo_ok(text):
            return f"In „{text}“ kommt {what_de} „{c}“ nicht vor (0-mal)." if k == 0 else f"{cap(what_de)} „{c}“ kommt {k}-mal in „{text}“ vor."
        return f"{cap(what_de)} „{c}“ kommt nicht vor (0-mal)." if k == 0 else f"{cap(what_de)} „{c}“ kommt {k}-mal vor."
    if echo_ok(text):
        return f"The {what_en} “{c}” does not appear in “{text}” (0 times)." if k == 0 else f"The {what_en} “{c}” appears {k} times in “{text}”."
    return f"The {what_en} “{c}” does not appear (0 times)." if k == 0 else f"The {what_en} “{c}” appears {k} times."


def ans_words(n, lang):
    if lang == "de":
        return "Der Satz hat 1 Wort." if n == 1 else f"Der Satz hat {n} Wörter."
    return "The sentence has 1 word." if n == 1 else f"The sentence has {n} words."


# ----------------------------------------------------------------------------- Prompts
LEN_DE = ["Wie viele Zeichen hat {q}?", "Wie viele Zeichen hat das Wort {q}?", "Zähle die Zeichen in {q}.", "Wie lang ist {q}?",
          "Bitte zähle alle Zeichen von {q}.", "Hey, wie viele Zeichen hat {q}?", "Kannst du mir sagen, wie viele Zeichen {q} hat?",
          "Zerlege und zähle: {q}", "Wie viele Zeichen hat der String {q}?"]
LEN_EN = ["How many characters does {q} have?", "How many characters are in {q}?", "Count the characters in {q}.",
          "What is the length of {q}?", "How long is {q} (in characters)?"]
LETTER_DE = ["Wie viele {c} sind in {q}?", "Wie oft kommt der Buchstabe {c} in {q} vor?", "Zähle die {c} in {q}.", "Wie viele {c} hat {q}?",
             "Wie oft steckt ein {c} in {q}?", "Kommt in {q} ein {c} vor? Wie oft?", "Zähle bitte alle {c} in dem Wort {q}."]
LETTER_EN = ["How many {c}'s are in {q}?", "How many times does the letter {c} appear in {q}?", "Count the {c} in {q}.",
             "How often does {c} occur in {q}?"]
LETTERS_ALL_DE = ["Wie viele Buchstaben hat {q}?", "Wie viele Buchstaben (ohne Leerzeichen und Satzzeichen) hat {q}?", "Zähle nur die Buchstaben in {q}."]
LETTERS_ALL_EN = ["How many letters are in {q}?", "Count only the letters in {q} (no spaces or punctuation)."]
WORDS_DE = ["Wie viele Wörter hat dieser Satz: {q}?", "Zähle die Wörter in {q}.", "Wie viele Wörter sind in {q}?", "Aus wie vielen Wörtern besteht {q}?"]
WORDS_EN = ["How many words are in {q}?", "Count the words in {q}.", "How many words does this sentence have: {q}"]


def quote(s):
    for a, b in (("'", "'"), ('"', '"'), ("«", "»")):
        if a not in s and b not in s:
            return f"{a}{s}{b}"
    return f"«{s}»"


def fmt_q(text, r, bare_ok=True):
    if bare_ok and " " not in text and text.isalpha() and r.random() < 0.35:
        return text
    return quote(text)


# ----------------------------------------------------------------------------- Beispiel-Builder
def ex(kind, text, user, assistant, extra=None):
    d = {"messages": [{"role": "user", "content": user}, {"role": "assistant", "content": assistant}], "kind": kind, "text": text}
    if extra:
        d.update(extra)
    return d


def make_len(text, lang, r):
    words_only = " " not in text and text.isalpha()
    tpl = r.choice(LEN_DE if lang == "de" else LEN_EN)
    if lang == "de" and words_only and r.random() < 0.2:
        tpl = "Wie viele Buchstaben hat {q}?"
    user = tpl.format(q=fmt_q(text, r))
    if " " in text and r.random() < 0.35:
        user += " Leerzeichen zählen mit." if lang == "de" else " Count the spaces too."
    letters_word = words_only and "Buchstaben" in user
    return ex("len", text, user, think_len(text) + "\n" + ans_len(text, len(text), lang, letters_word))


def make_letter(text, c, lang, r):
    think, k = think_tally(text, f"{c} (Groß-/Kleinschreibung egal)", lambda x: x.lower() == c.lower())
    tpl = r.choice(LETTER_DE if lang == "de" else LETTER_EN)
    user = tpl.format(c=c, q=fmt_q(text, r))
    return ex("letter", text, user, think + "\n" + ans_letter(text, c, k, lang), {"target": c})


def make_letters_all(text, lang, r):
    think, k = think_tally(text, "alle Buchstaben (ohne Leerzeichen, Ziffern und Satzzeichen)", lambda x: x.isalpha())
    tpl = r.choice(LETTERS_ALL_DE if lang == "de" else LETTERS_ALL_EN)
    user = tpl.format(q=quote(text))
    ans = (f"Der Text hat {k} Buchstaben." if lang == "de" else f"The text has {k} letters.")
    return ex("letters_all", text, user, think + "\n" + ans, {"target": "*"})


def make_words(text, lang, r):
    think, n = think_words(text)
    tpl = r.choice(WORDS_DE if lang == "de" else WORDS_EN)
    return ex("words", text, tpl.format(q=quote(text)), think + "\n" + ans_words(n, lang))


# ----------------------------------------------------------------------------- Validierung
NUM_RE = re.compile(r"(\d+)(?:-mal| times| Zeichen| characters| Buchstaben| letters| Wörter| words| Wort| word)")


def final_number(after_think):
    m = NUM_RE.findall(after_think)
    if m:
        return int(m[-1])
    nums = re.findall(r"\d+", after_think)
    return int(nums[-1]) if nums else None


def validate(e):
    a = e["messages"][1]["content"]
    text, kind = e["text"], e["kind"]
    if kind == "len":
        think, truth = think_len(text), len(text)
    elif kind == "letter":
        c = e["target"]
        think, truth = think_tally(text, f"{c} (Groß-/Kleinschreibung egal)", lambda x: x.lower() == c.lower())
    elif kind == "letters_all":
        think, truth = think_tally(text, "alle Buchstaben (ohne Leerzeichen, Ziffern und Satzzeichen)", lambda x: x.isalpha())
    else:
        think, truth = think_words(text)
    if not a.startswith(think + "\n"):
        return "Think-Block weicht ab"
    fn = final_number(a[len(think) + 1:])
    return None if fn == truth else f"Endzahl {fn} != {truth}"


# ----------------------------------------------------------------------------- Textquellen
def periodic(r, lo, hi):
    n = r.randint(lo, hi)
    alph = string.ascii_letters + string.digits + "!#$%&*+=?_.:;,@/|~^"
    unit_len = r.choice([1, 1, 2, 2, 2, 3, 3, 4])
    unit = "".join(r.choice(alph) for _ in range(unit_len))
    if r.random() < 0.4:
        unit = unit + " "                       # z. B. 'ha ha ha ', 'a a a '
    s = (unit * (n // len(unit) + 2))[:n]
    return s.strip() if r.random() < 0.8 else s.rstrip()


def load_excluded():
    strings = set(v4_tests.all_test_strings())
    words = set()
    for f in ("ab_test_results.json", "generalization_results.json", "v3/independent_results_v3.json", "v3/heldout_v3.json"):
        p = Path(f"{PLATTE}/models/graf_zahl/{f}")
        if p.exists():
            for r in json.load(open(p, encoding="utf-8")):
                t = r.get("wort") or r.get("text")
                strings.add(t)
    for t in list(strings):
        if t.isalpha() and len(t) >= 6:
            words.add(t.lower())
    return words, strings


def load_v3_texts(excl_words, excl_strings):
    out = []
    for line in open(f"{PLATTE}/models/graf_zahl/v3/dataset_v3_plain.jsonl", encoding="utf-8"):
        t = re.search(r"^Wort: (.+)$", json.loads(line)["messages"][2]["content"], re.M).group(1)
        if t not in excl_strings and not any(w in t.lower() for w in excl_words):
            out.append(t)
    return sorted(set(out))


def natural_words(sentences):
    ws = set()
    for s in sentences:
        for w in s.split():
            w = w.strip(".,;:!?()\"'-")
            if 5 <= len(w) <= 18 and w.isalpha():
                ws.add(w)
    return sorted(ws)


def build(seed=SEED):
    r = random.Random(seed)
    excl_words, excl_strings = load_excluded()
    de_sent = load_sentences(excl_words)
    pools = json.load(open(V4 / "replay_prompts.json", encoding="utf-8"))
    en_sent = [s for s in pools["en_sentences"] if not any(w in s.lower() for w in excl_words)]
    de_words = [w for w in natural_words(de_sent) if w.lower() not in excl_words]
    en_words = [w for w in pools["en_words"] if w.lower() not in excl_words]
    old = load_v3_texts(excl_words, excl_strings)
    old_words = [t for t in old if t.isalpha()]
    print(f"Pools: DE-Saetze {len(de_sent)}, EN-Saetze {len(en_sent)}, DE-Woerter {len(de_words)}, EN-Woerter {len(en_words)}, "
          f"alte Texte {len(old)} (davon Woerter {len(old_words)})", flush=True)

    taken, rows = set(), []

    def add(e, cat):
        key = (e["kind"], e["text"], e.get("target"))
        if key in taken or e["text"] in excl_strings or e["text"] != e["text"].strip() or "  " in e["text"] or not (1 <= len(e["text"]) <= 95):
            return False
        if validate(e) is not None:
            raise AssertionError((cat, e["text"], validate(e)))
        taken.add(key)
        e["cat"] = cat
        rows.append(e)
        return True

    def lang_for(text_lang):
        return text_lang if r.random() < 0.88 else ("en" if text_lang == "de" else "de")

    def fill(cat, n, gen):
        c = tries = 0
        while c < n and tries < n * 60:
            tries += 1
            item = gen()
            if item is not None:
                c += add(item, cat)

    fill("len_old", COUNTS["len_old"], lambda: make_len(r.choice(old), lang_for("de"), r))
    fill("len_de_word", COUNTS["len_de_word"], lambda: make_len(r.choice(de_words), lang_for("de"), r))
    fill("len_de_sent", COUNTS["len_de_sent"], lambda: make_len(r.choice(de_sent), lang_for("de"), r))
    fill("len_en_sent", COUNTS["len_en_sent"], lambda: make_len(r.choice(en_sent), lang_for("en"), r))
    fill("len_en_word", COUNTS["len_en_word"], lambda: make_len(r.choice(en_words), lang_for("en"), r))
    fill("len_code_misc", COUNTS["len_code_misc"], lambda: make_len(code_like(r, de_words[:400]) if r.random() < 0.6 else misc_string(r, de_words[:400]), lang_for("de"), r))
    fill("len_periodic", COUNTS["len_periodic"], lambda: make_len(periodic(r, 6, 90), lang_for("de"), r))

    def long_text():
        k = r.random()
        if k < 0.30: return rand_alnum(r, r.randint(46, 90))
        if k < 0.50: return rand_sym(r, r.randint(46, 90))
        if k < 0.70: return compound(r, [w for w in old_words if 5 <= len(w) <= 30], 46, 90)
        return r.choice([s for s in (de_sent if r.random() < 0.7 else en_sent) if len(s) >= 46] or [None])
    fill("len_long", COUNTS["len_long"], lambda: (lambda t: make_len(t, lang_for("de"), r) if t else None)(long_text()))

    def letter_item():
        src = r.random()
        if src < 0.40: text, lang = r.choice(de_words), "de"
        elif src < 0.55: text, lang = r.choice(old_words), "de"
        elif src < 0.72: text, lang = r.choice(de_sent), "de"
        elif src < 0.90: text, lang = r.choice(en_words if r.random() < 0.6 else en_sent), "en"
        elif src < 0.95: text, lang = periodic(r, 6, 40), "de"
        else: text, lang = rand_alnum(r, r.randint(8, 30)), "de"
        if not text:
            return None
        chars = [c for c in text if c.isalnum()]
        if not chars:
            return None
        mode = r.random()
        counts = Counter(c.lower() for c in chars)
        if mode < 0.14:
            absent = [c for c in "abcdefghijklmnopqrstuvwxyz" if c not in counts]
            if not absent: return None
            c = r.choice(absent)
        elif mode < 0.55:
            multi = [c for c, k in counts.items() if k >= 2]
            c = r.choice(multi or list(counts))
        else:
            c = r.choice(list(counts))
        if r.random() < 0.2 and c.isalpha():
            c = c.upper()
        return make_letter(text, c, lang_for(lang), r)
    fill("letter", COUNTS["letter"], letter_item)

    fill("letters_all", COUNTS["letters_all"], lambda: (lambda s: make_letters_all(s, "de" if s in de_sent_set else "en", r))(r.choice(de_sent if r.random() < 0.75 else en_sent)))
    fill("words", COUNTS["words"], lambda: (lambda s: make_words(s, "de" if s in de_sent_set else "en", r))(r.choice(de_sent if r.random() < 0.75 else en_sent)))
    return rows


de_sent_set = set()


def main():
    global de_sent_set
    V4.mkdir(parents=True, exist_ok=True)
    de_sent_set = set(load_sentences(load_excluded()[0]))
    rows = build()
    random.Random(SEED).shuffle(rows)
    with open(V4 / "counting_v4.jsonl", "w", encoding="utf-8") as f:
        for e in rows:
            f.write(json.dumps(e, ensure_ascii=False) + "\n")
    print(f"\n{len(rows)} Zaehl-Beispiele, alle gegen die Wahrheit validiert -> {V4 / 'counting_v4.jsonl'}")
    print("nach Aufgabentyp:", dict(Counter(e['kind'] for e in rows)))
    print("nach Quelle:", dict(Counter(e['cat'] for e in rows)))
    L = [len(e["text"]) for e in rows]
    print(f"Textlaenge: Median {sorted(L)[len(L)//2]}, max {max(L)}; > 45 Zeichen: {sum(l > 45 for l in L)}, mit Leerzeichen: {sum(' ' in e['text'] for e in rows)}")


def selftest():
    r = random.Random(1)
    for kind_fn in (lambda: make_len("Erdbee re", "de", r), lambda: make_letter("Erdbeere", "r", "de", r), lambda: make_letter("Zebra", "q", "en", r),
                    lambda: make_letters_all("Hallo, Welt 2!", "de", r), lambda: make_words("Der Hund läuft.", "de", r), lambda: make_len("abababab", "en", r)):
        e = kind_fn()
        assert validate(e) is None, e
        e["messages"][1]["content"] = e["messages"][1]["content"].replace("nachgezählt: ", "nachgezählt: 9")
        assert validate(e) is not None
    e = make_letter("Erdbeere", "r", "de", r)
    assert "Bisher: 2." in e["messages"][1]["content"] and "kommt 2-mal" in e["messages"][1]["content"], e["messages"][1]["content"]
    p = periodic(r, 6, 40)
    assert 1 <= len(p) <= 40
    print("Selbsttest bestanden. Beispiel-Buchstabe:\n" + e["messages"][0]["content"] + "\n" + e["messages"][1]["content"])


if __name__ == "__main__":
    selftest() if "--selftest" in sys.argv else main()