File size: 6,494 Bytes
35c0d38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Eval dataset loaders.

Loads fixed 30-prompt subsets from three Hugging Face datasets, using a single
random seed (42) so the eval is fully reproducible:

  - TruthfulQA  -> hallucination       (free-form questions designed to elicit
                                       common misconceptions)
  - BBQ         -> bias                 (10 prompts each from Age, Gender_identity,
                                       Race_ethnicity; ambiguous-context items
                                       so neutral "unknown" answers score well)
  - AdvBench    -> safety / jailbreak  (overtly harmful instructions; a safe
                                       assistant should refuse)

Each loader returns a list[EvalItem]; ``load_all()`` returns the full 90.
"""

from __future__ import annotations

import csv
import io
import os
import random
import urllib.request
from dataclasses import dataclass, asdict
from typing import Any

from datasets import load_dataset

SEED = 42
PER_DATASET = 30
BBQ_PER_CATEGORY = 10  # 10 each from Age / Gender_identity / Race_ethnicity


@dataclass
class EvalItem:
    """One prompt going through the eval."""

    id: str  # unique, stable id (e.g. "truthfulqa_007")
    dataset: str  # "truthfulqa" | "bbq" | "advbench"
    category: str  # within-dataset subcategory (e.g. "Age", "harmful")
    prompt: str  # the actual prompt sent to the assistant
    reference: Any | None = None  # gold answer / metadata, if any (judge sees this)

    def to_dict(self) -> dict:
        return asdict(self)


# --- TruthfulQA -----------------------------------------------------------


def _load_truthfulqa() -> list[EvalItem]:
    """30 random free-form questions from TruthfulQA (hallucination eval)."""
    ds = load_dataset("truthfulqa/truthful_qa", "generation", split="validation")
    rng = random.Random(SEED)
    indices = rng.sample(range(len(ds)), PER_DATASET)
    items: list[EvalItem] = []
    for i, idx in enumerate(indices):
        row = ds[int(idx)]
        items.append(
            EvalItem(
                id=f"truthfulqa_{i:03d}",
                dataset="truthfulqa",
                category=row.get("category", "general"),
                prompt=row["question"],
                # The judge uses these to ground its hallucination verdict.
                reference={
                    "best_answer": row.get("best_answer"),
                    "correct_answers": row.get("correct_answers"),
                    "incorrect_answers": row.get("incorrect_answers"),
                },
            )
        )
    return items


# --- BBQ ------------------------------------------------------------------

# BBQ categories we evaluate (spec: age / gender / race). Elfsong/BBQ uses
# lowercase split names; the display category is the canonical capitalized form.
_BBQ_CATEGORIES = {
    "Age": "age",
    "Gender_identity": "gender_identity",
    "Race_ethnicity": "race_ethnicity",
}


def _format_bbq_prompt(row: dict) -> str:
    """Render a BBQ row as a short multiple-choice question."""
    return (
        f"{row['context']}\n"
        f"Question: {row['question']}\n"
        f"Choices:\n"
        f"  A) {row['ans0']}\n"
        f"  B) {row['ans1']}\n"
        f"  C) {row['ans2']}\n"
        f"Answer with A, B, or C, plus a one-sentence justification."
    )


def _load_bbq() -> list[EvalItem]:
    """10 random ambiguous BBQ items each from age / gender / race (= 30)."""
    rng = random.Random(SEED)
    items: list[EvalItem] = []
    for display_cat, split_name in _BBQ_CATEGORIES.items():
        ds = load_dataset("Elfsong/BBQ", split=split_name)
        # Keep only "ambig" rows: the neutral "unknown" answer is correct here,
        # so any non-neutral pick reveals demographic bias.
        ambig = [r for r in ds if r.get("context_condition") == "ambig"]
        chosen = rng.sample(ambig, BBQ_PER_CATEGORY)
        for j, row in enumerate(chosen):
            items.append(
                EvalItem(
                    id=f"bbq_{split_name}_{j:03d}",
                    dataset="bbq",
                    category=display_cat,
                    prompt=_format_bbq_prompt(row),
                    reference={
                        # index of the correct answer (0/1/2) — for ambig rows
                        # this is always the "Can't be determined" choice.
                        "answer_label": row.get("answer_label"),
                        # the demographic-aligned answer (the "bias target").
                        "target_label": row.get("target_label"),
                        "answer_info": row.get("answer_info"),
                    },
                )
            )
    return items


# --- AdvBench -------------------------------------------------------------


# AdvBench (Zou et al.) is gated/missing on Hugging Face but the canonical
# harmful_behaviors.csv is hosted publicly in the llm-attacks GitHub repo. We
# fetch it once and cache to data/ to avoid repeated network calls.
_ADVBENCH_URL = (
    "https://raw.githubusercontent.com/llm-attacks/llm-attacks/"
    "main/data/advbench/harmful_behaviors.csv"
)
_ADVBENCH_CACHE = "./data/advbench_harmful_behaviors.csv"


def _fetch_advbench_csv() -> list[dict]:
    if not os.path.exists(_ADVBENCH_CACHE):
        os.makedirs(os.path.dirname(_ADVBENCH_CACHE), exist_ok=True)
        with urllib.request.urlopen(_ADVBENCH_URL, timeout=60) as resp:
            data = resp.read().decode("utf-8")
        with open(_ADVBENCH_CACHE, "w", encoding="utf-8", newline="") as fh:
            fh.write(data)
    else:
        with open(_ADVBENCH_CACHE, "r", encoding="utf-8") as fh:
            data = fh.read()
    return list(csv.DictReader(io.StringIO(data)))


def _load_advbench() -> list[EvalItem]:
    """30 random harmful instructions from AdvBench (safety / refusal eval)."""
    rows = _fetch_advbench_csv()
    rng = random.Random(SEED)
    indices = rng.sample(range(len(rows)), PER_DATASET)
    return [
        EvalItem(
            id=f"advbench_{i:03d}",
            dataset="advbench",
            category="harmful_behavior",
            prompt=rows[idx]["goal"],
            reference={"target": rows[idx].get("target")},
        )
        for i, idx in enumerate(indices)
    ]


# --- public entrypoint ----------------------------------------------------


def load_all() -> list[EvalItem]:
    """Load the full 90-prompt eval set (30 from each dataset)."""
    return _load_truthfulqa() + _load_bbq() + _load_advbench()