File size: 10,885 Bytes
96dc6b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Score Perch benchmark methods with TypeLLM's hosted boolean API.

Every method side is an independent request. Applicable bug/security questions
share that request; a semantic-lint pair uses its same ensure rule on both sides.
The JSONL predictions are append-only and can be resumed after interruption.
"""

from __future__ import annotations

import argparse
from concurrent.futures import ThreadPoolExecutor, as_completed
import hashlib
import importlib.util
import json
import math
import os
from pathlib import Path
import random
import sys
import threading
import time
from urllib.error import HTTPError, URLError
from urllib.request import Request, urlopen


MODEL = "typellm-latest"
API = "https://api.typellm.ai/v1/generate"
PROTOCOL = "typellm-boolean-v1"
LINT_PROTOCOL = "perch-semantic-ensure-v1"
DEFAULT_INTERVAL = 0.62  # Below the documented account cap of 100 requests/minute.


def module_at(name: str, path: Path):
    spec = importlib.util.spec_from_file_location(name, path)
    if spec is None or spec.loader is None:
        raise ValueError(f"Cannot import {path}")
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module


def probability(value: object) -> float:
    if isinstance(value, bool):
        raise ValueError("Boolean is not a probability")
    score = float(value)
    if not math.isfinite(score) or not 0 <= score <= 1:
        raise ValueError("Invalid probability")
    return score


class RateLimiter:
    def __init__(self, interval: float):
        self.interval = interval
        self.lock = threading.Lock()
        self.next_at = time.monotonic()

    def wait(self):
        with self.lock:
            now = time.monotonic()
            at = max(now, self.next_at)
            self.next_at = at + self.interval
        if at > now:
            time.sleep(at - now)

    def back_off(self, seconds: float):
        with self.lock:
            self.next_at = max(self.next_at, time.monotonic() + seconds)


def type_question(question: dict) -> dict:
    item = question["question"]
    true = str(item["criteria"]["true"]).strip()
    false = str(item["criteria"]["false"]).strip()
    return {
        "type": "boolean",
        "instructions": f'{item["instructions"].strip()}\nTrue: {true}\nFalse: {false}',
        "return_probabilities": True,
    }


def lint_question(ensure: str) -> dict:
    rule = " ".join(ensure.split())
    return {
        "type": "boolean",
        "instructions": f"Is `rule` true of the code below?\nRule: {rule}\nTrue: {rule}\nFalse: Not so: {rule}",
        "return_probabilities": True,
    }


def true_score(result: dict) -> float:
    values = result.get("probabilities")
    if not isinstance(values, dict) or "true" not in values or "false" not in values:
        raise ValueError("TypeLLM did not return both boolean probabilities")
    true, false = probability(values["true"]), probability(values["false"])
    if abs(true + false - 1) > 1e-5:
        raise ValueError("Boolean probabilities do not sum to one")
    return true


def invoke(key: str, state: dict, questions: dict, limiter: RateLimiter) -> dict:
    context = "Perch method context (JSON):\n" + json.dumps(state, ensure_ascii=False, separators=(",", ":"))
    payload = json.dumps({"model": MODEL, "context": context, "questions": questions,
                          "options": {"temperature": 0}, "timeout": 90}, ensure_ascii=False).encode()
    headers = {"Authorization": "Bearer " + key, "Content-Type": "application/json"}
    for attempt in range(8):
        limiter.wait()
        started = time.monotonic()
        request = Request(API, data=payload, headers=headers, method="POST")
        try:
            with urlopen(request, timeout=105) as response:
                data = json.load(response)
            if data.get("model") != MODEL or set(data.get("result", {})) != set(questions):
                raise ValueError("TypeLLM returned a different model or incomplete result")
            data["request_elapsed_s"] = round(time.monotonic() - started, 3)
            return data
        except HTTPError as exc:
            if exc.code not in (429, 502, 503, 504) or attempt == 7:
                raise RuntimeError(f"TypeLLM HTTP {exc.code}") from None
            header = exc.headers.get("Retry-After")
            try:
                delay = max(1.0, min(90.0, float(header)))
            except (ValueError, TypeError):
                delay = min(60.0, 2.0 ** attempt)
            limiter.back_off(delay)
            time.sleep(delay + random.random() * 0.2)
        except URLError:
            if attempt == 7:
                raise RuntimeError("TypeLLM transport failed after retries") from None
            time.sleep(min(30, 2 ** attempt))
    raise AssertionError("unreachable")


def load_pairs(path: Path, track: str):
    raw = path.read_bytes()
    pairs = [json.loads(line) for line in raw.splitlines() if line]
    if not pairs:
        raise ValueError("Empty benchmark")
    seen = set()
    for pair in pairs:
        pair_id = pair["pair_id"]
        if pair_id in seen:
            raise ValueError("Duplicate pair ID")
        seen.add(pair_id)
        if track != "lint" and pair["domain"] != track:
            raise ValueError("Wrong benchmark domain")
        if track == "lint" and not pair.get("rule", {}).get("ensure"):
            raise ValueError("Missing semantic lint rule")
    return pairs, hashlib.sha256(raw).hexdigest()


def saved_predictions(path: Path, pairs: list[dict], digest: str, track: str, pack: dict | None):
    expected = {(pair["pair_id"], role) for pair in pairs for role in ("before", "after")}
    rows = {}
    if not path.exists():
        return rows
    for line in path.open():
        row = json.loads(line)
        key = row["pair_id"], row["role"]
        if key not in expected or key in rows or row["dataset_sha256"] != digest:
            raise ValueError("Unexpected or duplicate saved prediction")
        if track == "lint":
            if row["protocol"] != LINT_PROTOCOL or row["model"] != MODEL:
                raise ValueError("Incompatible saved lint prediction")
            if abs(probability(row["score"]) + probability(row["rule_true_probability"]) - 1) > 1e-8:
                raise ValueError("Invalid saved lint score")
        else:
            if (row["protocol_id"] != pack["protocol_id"] or row.get("interface_id") != PROTOCOL
                    or row["model_revision"] != MODEL
                    or row["pack_sha256"] != pack["pack_sha256"]):
                raise ValueError("Incompatible saved scan prediction")
        rows[key] = row
    return rows


def score_one(pair: dict, role: str, track: str, pack: dict | None, core,
              key: str, digest: str, limiter: RateLimiter, threshold: float) -> dict:
    state = pair[role]["state"]
    if track == "lint":
        questions = {"rule_true": lint_question(pair["rule"]["ensure"])}
    else:
        selected = core.detection_questions(core.scan_questions(pack, track, pair["language"]), track)
        questions = {question["name"]: type_question(question) for question in selected}
        if not questions:
            raise ValueError("No applicable detection questions")
    response = invoke(key, state, questions, limiter)
    usage = response.get("usage") or {}
    if track == "lint":
        true = true_score(response["result"]["rule_true"])
        return {"pair_id": pair["pair_id"], "role": role, "domain": pair["domain"],
                "model": MODEL, "protocol": LINT_PROTOCOL, "dataset_sha256": digest,
                "rule_true_probability": true, "score": 1 - true,
                "usage": usage, "elapsed_s": response["request_elapsed_s"]}
    scores = {name: core.raised_score(question, true_score(response["result"][name]))
              for question in selected for name in (question["name"],)}
    score, alert = core.scan_result(scores, selected, track, threshold)
    return {"pair_id": pair["pair_id"], "role": role, "domain": track,
            "model_id": "typellm", "model_revision": MODEL,
            "protocol_id": pack["protocol_id"], "interface_id": PROTOCOL,
            "pack_sha256": pack["pack_sha256"],
            "dataset_sha256": digest, "question_scores": scores,
            "score": score, "alert": alert, "usage": usage,
            "elapsed_s": response["request_elapsed_s"]}


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--track", choices=("bug", "security", "lint"), required=True)
    parser.add_argument("--dataset", type=Path, required=True)
    parser.add_argument("--predictions", type=Path, required=True)
    parser.add_argument("--question-pack", type=Path, default=Path(__file__).with_name("question-pack.json"))
    parser.add_argument("--benchmark-script", type=Path, default=Path(__file__).with_name("benchmark.py"))
    parser.add_argument("--threshold", type=float, default=0.5)
    parser.add_argument("--workers", type=int, default=8)
    parser.add_argument("--interval", type=float, default=DEFAULT_INTERVAL)
    parser.add_argument("--limit", type=int, default=0)
    args = parser.parse_args()
    if not 0 <= args.threshold <= 1 or args.workers < 1 or args.interval < 0.6 or args.limit < 0:
        parser.error("Invalid threshold, workers, interval, or limit")
    key = os.environ.get("TYPELLM_API_KEY")
    if not key:
        raise RuntimeError("Set TYPELLM_API_KEY")
    core = module_at("perch_benchmark", args.benchmark_script) if args.track != "lint" else None
    pack = core.load_pack(args.question_pack) if core else None
    pairs, digest = load_pairs(args.dataset, args.track)
    existing = saved_predictions(args.predictions, pairs, digest, args.track, pack)
    pending = [(pair, role) for pair in pairs for role in ("before", "after")
               if (pair["pair_id"], role) not in existing]
    if args.limit:
        pending = pending[:args.limit]
    if not pending:
        print(f"{args.track}: already complete", flush=True)
        return
    args.predictions.parent.mkdir(parents=True, exist_ok=True)
    limiter = RateLimiter(args.interval)
    with args.predictions.open("a") as output, ThreadPoolExecutor(max_workers=args.workers) as pool:
        futures = [pool.submit(score_one, pair, role, args.track, pack, core, key, digest,
                               limiter, args.threshold)
                   for pair, role in pending]
        for index, future in enumerate(as_completed(futures), 1):
            row = future.result()
            output.write(json.dumps(row, sort_keys=True) + "\n")
            output.flush()
            if index % 100 == 0 or index == len(pending):
                print(f"{args.track}: {len(existing) + index}/{2 * len(pairs)} sides", flush=True)


if __name__ == "__main__":
    main()