File size: 10,202 Bytes
7b4b748
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Load public safety/factuality benchmarks from Hugging Face datasets."""

from __future__ import annotations

import logging
import random
from collections.abc import Callable

from evaluation.prompts import EvalMetric, EvalPrompt

logger = logging.getLogger(__name__)

BenchmarkLoader = Callable[[int, int], list[EvalPrompt]]


def _sample_rows(rows: list[EvalPrompt], limit: int, seed: int) -> list[EvalPrompt]:
    if limit <= 0 or limit >= len(rows):
        return rows
    rng = random.Random(seed)
    return rng.sample(rows, limit)


def _get_row_field(row: dict, *keys: str):
    for key in keys:
        value = row.get(key)
        if value is not None and str(value).strip():
            return value
    return None


def _truthfulqa_reference(row: dict) -> str | None:
    best = _get_row_field(row, "best_answer", "Best Answer")
    correct = _get_row_field(row, "correct_answers", "Correct Answers")
    if isinstance(correct, str):
        parts = [part.strip() for part in correct.split(";") if part.strip()]
        correct = parts
    elif not correct:
        correct = []
    if best:
        return str(best).strip()
    if correct:
        return str(correct[0]).strip()
    return None


def load_truthfulqa(limit: int = 5, seed: int = 42) -> list[EvalPrompt]:
    """TruthfulQA (Lin et al.) — public factuality / hallucination benchmark."""
    from datasets import load_dataset

    dataset_candidates = [
        ("domenicrosati/TruthfulQA", None, "train"),
        ("truthfulqa/truthful_qa", "generation", "validation"),
        ("leibni/truthful_qa", "generation", "validation"),
    ]

    for dataset_name, config, split in dataset_candidates:
        try:
            if config:
                ds = load_dataset(dataset_name, config, split=split)
            else:
                ds = load_dataset(dataset_name, split=split)

            prompts: list[EvalPrompt] = []
            for idx, row in enumerate(ds):
                question = _get_row_field(row, "question", "Question")
                if not question:
                    continue
                prompts.append(
                    EvalPrompt(
                        id=f"truthfulqa_{idx}",
                        metric="hallucination",
                        category="factual",
                        prompt=str(question).strip(),
                        reference=_truthfulqa_reference(row),
                        expect_refusal=False,
                        notes=f"TruthfulQA from {dataset_name}",
                        source="public",
                        benchmark="truthfulqa",
                    )
                )

            sampled = _sample_rows(prompts, limit, seed)
            if sampled:
                logger.info("Loaded TruthfulQA from %s", dataset_name)
                return sampled
        except Exception as exc:
            logger.warning("Could not load TruthfulQA from %s: %s", dataset_name, exc)

    logger.warning("All TruthfulQA sources failed; returning empty set")
    return []


def load_real_toxicity_prompts(limit: int = 5, seed: int = 42) -> list[EvalPrompt]:
    """RealToxicityPrompts (Gehman et al.) — toxicity / harmful generation benchmark."""
    from datasets import load_dataset

    ds = load_dataset("allenai/real-toxicity-prompts", split="train")
    prompts: list[EvalPrompt] = []

    for idx, row in enumerate(ds):
        text = row.get("prompt", {}).get("text") or row.get("text") or ""
        text = str(text).strip()
        if not text:
            continue
        prompts.append(
            EvalPrompt(
                id=f"rtp_{idx}",
                metric="bias_harm",
                category="bias",
                prompt=text,
                expect_refusal=True,
                notes="RealToxicityPrompts",
                source="public",
                benchmark="real_toxicity_prompts",
            )
        )

    return _sample_rows(prompts, limit, seed)


def _extract_prompt_text(row: dict) -> str:
    for key in (
        "prompt",
        "goal",
        "question",
        "instruction",
        "text",
        "adversarial",
        "vanilla",
        "jailbreak_query",
        "behavior",
    ):
        value = row.get(key)
        if value and str(value).strip():
            return str(value).strip()
    return ""


def _rows_to_adversarial_prompts(
    rows,
    *,
    benchmark_name: str,
    source_note: str,
    id_prefix: str,
) -> list[EvalPrompt]:
    prompts: list[EvalPrompt] = []
    for idx, row in enumerate(rows):
        text = _extract_prompt_text(row)
        if not text:
            continue
        prompts.append(
            EvalPrompt(
                id=f"{id_prefix}_{idx}",
                metric="content_safety",
                category="adversarial",
                prompt=text,
                expect_refusal=True,
                notes=source_note,
                source="public",
                benchmark=benchmark_name,
            )
        )
    return prompts


def load_simple_safety_tests(limit: int = 5, seed: int = 42) -> list[EvalPrompt]:
    """Safety/jailbreak benchmarks — SimpleSafetyTests + open fallbacks."""
    from datasets import load_dataset

    def load_sst_instruct() -> list[EvalPrompt]:
        ds = load_dataset("walledai/SimpleSafetyTests", split="instruct")
        return _rows_to_adversarial_prompts(
            ds,
            benchmark_name="simple_safety_tests",
            source_note="walledai/SimpleSafetyTests (instruct)",
            id_prefix="sst",
        )

    def load_jailbreak_wild() -> list[EvalPrompt]:
        ds = load_dataset(
            "TrustAIRLab/in-the-wild-jailbreak-prompts",
            "jailbreak_2023_05_07",
            split="train",
        )
        return _rows_to_adversarial_prompts(
            ds,
            benchmark_name="simple_safety_tests",
            source_note="TrustAIRLab/in-the-wild-jailbreak-prompts",
            id_prefix="jailbreak",
        )

    def load_forbidden_questions() -> list[EvalPrompt]:
        ds = load_dataset("TrustAIRLab/forbidden_question_set", split="train")
        return _rows_to_adversarial_prompts(
            ds,
            benchmark_name="simple_safety_tests",
            source_note="TrustAIRLab/forbidden_question_set",
            id_prefix="forbidden",
        )

    for name, loader in [
        ("SimpleSafetyTests", load_sst_instruct),
        ("in-the-wild-jailbreak-prompts", load_jailbreak_wild),
        ("forbidden_question_set", load_forbidden_questions),
    ]:
        try:
            prompts = loader()
            sampled = _sample_rows(prompts, limit, seed)
            if sampled:
                logger.info("Loaded safety benchmark from %s", name)
                return sampled
        except Exception as exc:
            logger.warning("Could not load safety benchmark %s: %s", name, exc)

    logger.warning("All safety benchmark sources failed; returning empty set")
    return []


def _load_bbq_subset(dataset_name: str, config: str | None, limit: int, seed: int) -> list[EvalPrompt]:
    from datasets import load_dataset

    if config:
        ds = load_dataset(dataset_name, config, split="test")
    else:
        ds = load_dataset(dataset_name, split="test")

    prompts: list[EvalPrompt] = []
    for idx, row in enumerate(ds):
        context = str(row.get("context", "")).strip()
        question = str(row.get("question", "")).strip()
        if not question:
            continue
        prompt = f"{context}\n\n{question}".strip() if context else question
        prompts.append(
            EvalPrompt(
                id=f"bbq_{idx}",
                metric="bias_harm",
                category="bias",
                prompt=prompt,
                expect_refusal=False,
                notes=f"BBQ subset from {dataset_name}",
                source="public",
                benchmark="bbq",
            )
        )
    return _sample_rows(prompts, limit, seed)


def load_bbq(limit: int = 5, seed: int = 42) -> list[EvalPrompt]:
    """BBQ (Bias Benchmark for QA, Parrish et al.) — social bias benchmark."""
    candidates = [
        ("oskarvanderwal/bbq", "Age"),
        ("Heegyu/bbq", None),
    ]
    last_error: Exception | None = None

    for dataset_name, config in candidates:
        try:
            return _load_bbq_subset(dataset_name, config, limit, seed)
        except Exception as exc:
            last_error = exc
            logger.warning("Could not load BBQ subset %s: %s", dataset_name, exc)

    raise RuntimeError("Failed to load BBQ benchmark") from last_error


PUBLIC_BENCHMARKS: dict[str, BenchmarkLoader] = {
    "truthfulqa": load_truthfulqa,
    "real_toxicity_prompts": load_real_toxicity_prompts,
    "simple_safety_tests": load_simple_safety_tests,
    "bbq": load_bbq,
}


def load_public_benchmarks(
    names: list[str] | None = None,
    samples_per_benchmark: int = 5,
    seed: int = 42,
) -> list[EvalPrompt]:
    if names and "all" in names:
        names = list(PUBLIC_BENCHMARKS.keys())
    selected = names or list(PUBLIC_BENCHMARKS.keys())
    prompts: list[EvalPrompt] = []
    failures: list[str] = []

    for name in selected:
        if name not in PUBLIC_BENCHMARKS:
            raise ValueError(
                f"Unknown benchmark '{name}'. Available: {', '.join(PUBLIC_BENCHMARKS)}"
            )
        loader = PUBLIC_BENCHMARKS[name]
        try:
            loaded = loader(samples_per_benchmark, seed)
            if not loaded:
                failures.append(name)
                logger.warning("No prompts loaded for public benchmark '%s'", name)
                continue
            prompts.extend(loaded)
            logger.info("Loaded %d prompts from public benchmark '%s'", len(loaded), name)
        except Exception as exc:
            failures.append(name)
            logger.warning("Failed to load public benchmark '%s': %s", name, exc)

    if failures:
        logger.warning("Skipped failed benchmarks: %s", ", ".join(failures))
    if not prompts:
        raise RuntimeError(
            f"Could not load any public benchmarks. Failures: {', '.join(failures)}"
        )

    return prompts