File size: 17,017 Bytes
ba82887
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f3f832d
ba82887
 
 
 
 
f3f832d
ba82887
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
"""NNAI-Bench v0.1: deterministic synthetic tasks, scoring and CPU baselines.

Python 3.10+, standard library only. No CAEL dependencies or neural models.
"""
from __future__ import annotations

import argparse
from collections import Counter, defaultdict
import hashlib
import json
import math
from pathlib import Path
import random
import re
import time

VERSION = "nnai-bench-0.1"
ROOT = Path(__file__).resolve().parent
DATASET = ROOT / "dataset" if (ROOT / "dataset").is_dir() else ROOT
ARTIFACTS = ROOT / "space" if (ROOT / "space").is_dir() else ROOT
CATEGORIES = ("arithmetic", "logic", "ordering", "state_tracking", "strings", "tables")
LETTERS = "ABCD"
SPLITS = {"train": 100, "validation": 50, "test": 150}


def dump(path, value):
    path = Path(path)
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8", newline="\n")


def write_rows(path, rows):
    path = Path(path)
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows), encoding="utf-8", newline="\n")


def read_rows(path):
    return [json.loads(line) for line in Path(path).read_text(encoding="utf-8").splitlines() if line.strip()]


def numeric_choices(answer, rng):
    options = {str(answer)}
    while len(options) < 4:
        options.add(str(answer + rng.choice((-1, 1)) * rng.randint(1, 19)))
    return sorted(options)


def make_task(category, tier, rng, tag):
    """Return a prompt, four distinct options, answer and independently auditable program."""
    size = (3, 5, 8)[tier - 1]
    if category == "arithmetic":
        a, b, c = (rng.randint(3, 99) for _ in range(3))
        if tier == 1:
            answer, expression = a + b, f"{a} + {b}"
        elif tier == 2:
            answer, expression = a * b - c, f"({a} * {b}) - {c}"
        else:
            answer, expression = (a + b) * c - a, f"(({a} + {b}) * {c}) - {a}"
        prompt = rng.choice((f"Calculate {expression}. What is the result?", f"Using standard arithmetic, evaluate {expression}."))
        return prompt, numeric_choices(answer, rng), str(answer), {"expression": expression, "value": answer}
    if category == "logic":
        predicates = [f"kind_{tag}_{i}" for i in range(size + 2)]
        entity = f"object_{tag}"
        edges = list(zip(predicates[:size], predicates[1:size + 1]))
        mode = rng.randrange(3)
        target = predicates[size] if mode != 2 else predicates[-1]
        facts = [f"{entity} is a {predicates[0]}."]
        facts += [f"Every {a} is a {b}." for a, b in edges]
        negatives = [target] if mode == 1 else []
        if negatives:
            # Make negative-only cases consistent: remove the final implication.
            edges = edges[:-1]
            facts = [facts[0]] + [f"Every {a} is a {b}." for a, b in edges]
            facts.append(f"{entity} is not a {target}.")
        rng.shuffle(facts)
        answer = ("true", "false", "unknown")[mode]
        prompt = "\n".join(facts) + f"\nUnder open-world logic, is '{entity} is a {target}' true, false, or unknown? No fact implies its converse."
        return prompt, ["true", "false", "unknown", "inconsistent"], answer, {"start": predicates[0], "edges": edges, "negatives": negatives, "target": target}
    if category == "ordering":
        objects = [f"item_{tag}_{i}" for i in range(size + 1)]
        order = rng.sample(objects, len(objects))
        rules = [f"{a} comes before {b}." for a, b in zip(order, order[1:])]
        rng.shuffle(rules)
        if tier == 1:
            index, question = 0, "Which item comes first?"
        elif tier == 2:
            index, question = len(order) - 1, "Which item comes last?"
        else:
            index = rng.randint(1, len(order) - 2)
            question = f"Which item comes immediately after {order[index - 1]}?"
        answer = order[index]
        options = [answer] + rng.sample([x for x in order if x != answer], 3)
        return "\n".join(rules + [question]), options, answer, {"order": order, "index": index}
    if category == "state_tracking":
        people = [f"person_{tag}_{i}" for i in range(4)]
        rooms = [f"room_{tag}_{i}" for i in range(4)]
        events = []
        for person, room in zip(people, rng.sample(rooms, 4)):
            events.append([person, room])
        for _ in range(size * 2):
            events.append([rng.choice(people), rng.choice(rooms)])
        target = rng.choice(people)
        state = {p: r for p, r in events}
        prompt = "Events are listed from earliest to latest.\n" + "\n".join(f"{p} went to {r}." for p, r in events)
        prompt += f"\nWhere is {target} now?"
        return prompt, rooms, state[target], {"events": events, "target": target}
    if category == "strings":
        text = "".join(rng.choice("abcdefghjkmnpqrstuvwxyz") for _ in range(size + 2))
        if tier == 1:
            answer, operation = text[::-1], "reverse"
            question = f'Reverse the string "{text}".'
        elif tier == 2:
            answer, operation = text[1::2], "even_positions"
            question = f'From "{text}", keep characters at positions 2, 4, 6, and so on (1-based).'
        else:
            k = rng.randint(1, len(text) - 1)
            answer, operation = (text[k:] + text[:k])[::-1], "rotate_reverse"
            question = f'Rotate "{text}" left by {k} characters, then reverse the result.'
        options = {answer, text, text[::-1], text[::2], text[1::2]}
        while len(options) < 4:
            options.add(text + rng.choice("xyz"))
        options = [answer] + rng.sample(sorted(options - {answer}), 3)
        return question, options, answer, {"text": text, "operation": operation, "k": k if tier == 3 else 0}
    if category == "tables":
        records = [{"item": f"row_{tag}_{i}", "group": rng.choice(("red", "blue")), "units": rng.randint(1, 40), "price": rng.randint(2, 20)} for i in range(size + 1)]
        group = rng.choice(("red", "blue"))
        selected = [r for r in records if r["group"] == group]
        if tier == 1:
            answer, operation = len(selected), "count"
            question = f"How many rows have group {group}?"
        elif tier == 2:
            answer, operation = sum(r["units"] for r in selected), "sum"
            question = f"What is the total units for group {group}?"
        else:
            answer, operation = sum(r["units"] * r["price"] for r in selected), "weighted_sum"
            question = f"What is the sum of units * price for group {group}?"
        table = "item | group | units | price\n" + "\n".join(f'{r["item"]} | {r["group"]} | {r["units"]} | {r["price"]}' for r in records)
        return table + "\n" + question, numeric_choices(answer, rng), str(answer), {"records": records, "group": group, "operation": operation}
    raise ValueError(category)


def audit_answer(category, program):
    """A separate ground-truth checker; never a model input or ranked solver."""
    p = program
    if category == "arithmetic":
        # Only generator-owned numeric expressions; never evaluate submitted code.
        import ast
        import operator
        operators = {ast.Add: operator.add, ast.Sub: operator.sub, ast.Mult: operator.mul}
        def calc(node):
            if isinstance(node, ast.Constant) and type(node.value) is int:
                return node.value
            if isinstance(node, ast.BinOp) and type(node.op) in operators:
                return operators[type(node.op)](calc(node.left), calc(node.right))
            raise ValueError("Unsupported expression")
        return str(calc(ast.parse(p["expression"], mode="eval").body))
    if category == "logic":
        known = {p["start"]}
        while True:
            expanded = known | {b for a, b in p["edges"] if a in known}
            if expanded == known:
                break
            known = expanded
        pos, neg = p["target"] in known, p["target"] in p["negatives"]
        return "inconsistent" if pos and neg else "true" if pos else "false" if neg else "unknown"
    if category == "ordering":
        return p["order"][p["index"]]
    if category == "state_tracking":
        return next(room for person, room in reversed(p["events"]) if person == p["target"])
    if category == "strings":
        text = p["text"]
        if p["operation"] == "reverse":
            return "".join(reversed(text))
        if p["operation"] == "even_positions":
            return "".join(c for i, c in enumerate(text, 1) if i % 2 == 0)
        return "".join(reversed(text[p["k"]:] + text[:p["k"]]))
    selected = [r for r in p["records"] if r["group"] == p["group"]]
    if p["operation"] == "count":
        return str(len(selected))
    if p["operation"] == "sum":
        return str(sum(r["units"] for r in selected))
    return str(sum(r["units"] * r["price"] for r in selected))


def generate():
    output = DATASET
    seen = set()
    audits = []
    for split_index, (split, count) in enumerate(SPLITS.items()):
        rng = random.Random(20261018 + split_index)
        rows = []
        for category in CATEGORIES:
            positions = [(len(rows) + i) % 4 for i in range(count)]
            rng.shuffle(positions)
            for i in range(count):
                tier = i % 3 + 1
                while True:
                    tag = f"{split[0]}{rng.getrandbits(28):07x}"
                    prompt, choices, answer, program = make_task(category, tier, rng, tag)
                    if prompt not in seen:
                        seen.add(prompt)
                        break
                assert answer == audit_answer(category, program)
                wrong = sorted(set(choices) - {answer})
                rng.shuffle(wrong)
                label = positions[i]
                wrong.insert(label, answer)
                row = {"id": f"{split}-{category}-{i:04d}", "split": split, "category": category,
                       "difficulty": tier, "prompt": prompt, "choices": wrong, "answer": LETTERS[label]}
                rows.append(row)
                audits.append({"id": row["id"], "category": category, "program": program, "answer_text": answer})
        rng.shuffle(rows)
        write_rows(output / "data" / f"{split}.jsonl", rows)
        questions = [{k: v for k, v in r.items() if k != "answer"} for r in rows]
        write_rows(output / "questions" / f"{split}.jsonl", questions)
        if split == "test" and (ROOT / "space").is_dir():
            dump(ROOT / "space" / "test.json", rows)
    write_rows(output / "audit.jsonl", audits)
    checksums = {str(p.relative_to(output)).replace("\\", "/"): hashlib.sha256(p.read_bytes()).hexdigest()
                 for p in sorted(output.rglob("*.jsonl"))}
    manifest = {"benchmark_version": VERSION, "seed": 20261018,
                "split_sizes": {s: n * len(CATEGORIES) for s, n in SPLITS.items()},
                "categories": list(CATEGORIES), "sha256": checksums}
    dump(output / "manifest.json", manifest)
    if (ROOT / "space").is_dir():
        dump(ROOT / "space" / "manifest.json", manifest)
    dump(output / "example_predictions.json", [{"id": r["id"], "prediction": "ABSTAIN"} for r in read_rows(output / "data/test.jsonl")])
    print(json.dumps(manifest, indent=2))


def normalise_predictions(predictions, questions):
    if not isinstance(predictions, list):
        raise ValueError("Predictions must be a JSON array of {id, prediction} objects.")
    expected = {r["id"] for r in questions}
    got = {}
    for row in predictions:
        if not isinstance(row, dict) or set(row) != {"id", "prediction"}:
            raise ValueError("Each prediction must contain exactly id and prediction.")
        ident, pred = row["id"], row["prediction"]
        if not isinstance(ident, str) or ident not in expected or ident in got:
            raise ValueError(f"Unknown or duplicate ID: {ident!r}")
        if not isinstance(pred, str) or pred not in (*LETTERS, "ABSTAIN"):
            raise ValueError(f"Invalid prediction for {ident}; use A, B, C, D or ABSTAIN.")
        got[ident] = pred
    if set(got) != expected:
        raise ValueError(f"Expected {len(expected)} unique IDs, received {len(got)}. Use ABSTAIN for unanswered items.")
    return got


def score(predictions, questions):
    got = normalise_predictions(predictions, questions)
    counts = defaultdict(lambda: {"correct": 0, "answered": 0, "total": 0})
    for row in questions:
        c = counts[row["category"]]
        c["total"] += 1
        c["answered"] += got[row["id"]] != "ABSTAIN"
        c["correct"] += got[row["id"]] == row["answer"]
    categories = {k: {**v, "accuracy": 100 * v["correct"] / v["total"]} for k, v in sorted(counts.items())}
    total = len(questions)
    correct = sum(v["correct"] for v in counts.values())
    answered = sum(v["answered"] for v in counts.values())
    return {"benchmark_version": VERSION, "macro_accuracy": sum(v["accuracy"] for v in categories.values()) / len(categories),
            "accuracy": 100 * correct / total, "coverage": 100 * answered / total,
            "answered_accuracy": 100 * correct / answered if answered else 0,
            "correct": correct, "total": total, "categories": categories}


def tokens(row):
    return re.findall(r"[a-z]+|\d+", row["prompt"].lower())


def baseline_predictions(kind, train, test):
    rng = random.Random(314159)
    if kind == "seeded-random":
        return [{"id": r["id"], "prediction": rng.choice(LETTERS)} for r in test]
    if kind == "train-majority":
        # Deterministic tie-break; no test labels or test-derived label frequencies.
        counts = Counter(r["answer"] for r in train)
        best = max(LETTERS, key=lambda x: (counts[x], -LETTERS.index(x)))
        return [{"id": r["id"], "prediction": best} for r in test]
    # Unigram BM25 nearest training prompt, returning the neighbour's answer letter.
    # Deliberately simple retrieval baseline, without embeddings or test labels.
    docs = [Counter(tokens(r)) for r in train]
    df = Counter(t for d in docs for t in d)
    average = sum(sum(d.values()) for d in docs) / len(docs)
    index = defaultdict(list)
    for i, d in enumerate(docs):
        for term, tf in d.items():
            idf = math.log(1 + (len(docs) - df[term] + .5) / (df[term] + .5))
            weight = idf * tf * 2.5 / (tf + 1.5 * (.25 + .75 * sum(d.values()) / average))
            index[term].append((i, weight))
    predictions = []
    for row in test:
        values = defaultdict(float)
        for term in sorted(set(tokens(row))):
            for i, weight in index[term]:
                values[i] += weight
        best = max(values, key=lambda i: (values[i], -i)) if values else 0
        predictions.append({"id": row["id"], "prediction": train[best]["answer"]})
    return predictions


def baselines():
    train = read_rows(DATASET / "data/train.jsonl")
    labelled = read_rows(DATASET / "data/test.jsonl")
    inputs = [{k: v for k, v in r.items() if k != "answer"} for r in labelled]
    entries = []
    for kind in ("seeded-random", "train-majority", "bm25-nearest-training"):
        started = time.perf_counter()
        predictions = baseline_predictions(kind, train, inputs)
        elapsed = time.perf_counter() - started
        report = score(predictions, labelled)
        dump(ARTIFACTS / f"baselines/{kind}.predictions.json", predictions)
        dump(ARTIFACTS / f"baselines/{kind}.report.json", report)
        entries.append({"id": kind, "name": kind.replace("-", " ").title(), "kind": "baseline",
                        "status": "reproduced", "architecture": "BM25" if kind.startswith("bm25") else "statistical control",
                        "track": "closed-book", "training_provenance": "no neural assistance",
                        "source_url": "https://huggingface.co/datasets/vynly/nnai-bench/blob/main/benchmark.py",
                        "revision": VERSION, "runtime_seconds": round(elapsed, 5), "score": report})
    dump(ARTIFACTS / "leaderboard.json", {"benchmark_version": VERSION, "entries": entries})
    print(json.dumps([{r["name"]: r["score"]["macro_accuracy"]} for r in entries]))


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    sub = parser.add_subparsers(dest="command", required=True)
    sub.add_parser("generate")
    sub.add_parser("baselines")
    ev = sub.add_parser("score")
    ev.add_argument("predictions")
    ev.add_argument("--questions", default=str(DATASET / "data/test.jsonl"))
    ev.add_argument("--output")
    args = parser.parse_args()
    if args.command == "generate":
        generate()
    elif args.command == "baselines":
        baselines()
    else:
        report = score(json.loads(Path(args.predictions).read_text(encoding="utf-8")), read_rows(args.questions))
        if args.output:
            dump(args.output, report)
        print(json.dumps(report, indent=2))


if __name__ == "__main__":
    main()