Hugging Face's logo

fzzhang
/
svd-code

Model card Files
xet
 Community
svd-code
File size: 4,777 Bytes
58258b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Realistic science-prompt fixtures used by test_invariants.py.

The list `REALISTIC_SCIENCE_PROMPTS` is hand-crafted to contain four classes
of items so tests can pin exact post-stage counts:

  Class A: byte-identical duplicates (3 copies)
    -> caught by Stage 1 (caller's exact-prompt reservoir dedup)

  Class B: formatting variants of one prompt (3 prompts: extra whitespace,
           upper-case differences, trailing newline)
    -> NOT caught by exact dedup; caught by Stage 2 (MinHash with
       normalize=True)

  Class C: near-duplicate paraphrases of one prompt (3 prompts: each is the
           same long base text with one single-word edit)
    -> NOT caught by exact or formatting normalization; caught by Stage 2
       (MinHash with threshold=0.8)

  Class D: 7 distinct singleton prompts (math / biology / chemistry / physics)
    -> none of the dedup stages should touch these

Total = 3 + 3 + 3 + 7 = 16 prompts.
After exact dedup:    1 + 3 + 3 + 7 = 14
After MinHash:        1 + 1 + 1 + 7 = 10

Class C uses a ~140-word base text so that single-word edits keep Jaccard
robustly above MinHash's threshold of 0.8. With n=140 tokens (≈136 5-grams)
and a single-word edit affecting 5 5-grams, Jaccard = 131/141 ≈ 0.929 — well
into LSH's high-detection regime, so the test is deterministic.
"""

from __future__ import annotations


# ── Class A: byte-identical ─────────────────────────────────────────────────
_EXACT_DUP = (
    "Solve the equation x squared minus four equals zero. Show all steps "
    "clearly."
)

# ── Class B: formatting variants of one prompt ──────────────────────────────
_FMT_BASE = "Calculate the derivative of f(x) = x^3 + 2x with respect to x."
_FORMATTING_VARIANTS = [
    _FMT_BASE,
    "  Calculate  the  derivative  of  f(x)  =  x^3  +  2x  with  respect  to  x.  ",
    "Calculate the DERIVATIVE of f(x) = x^3 + 2x with respect to x.\n",
]

# ── Class C: near-duplicate paraphrases (single-word edits over ~80 words) ──
_NEAR_DUP_BASE = (
    "Explain in detail the process of photosynthesis in green plants. "
    "Describe step by step how light energy is captured by chlorophyll "
    "molecules in the thylakoid membrane, how electrons are transported "
    "through photosystems II and I, how ATP and NADPH are generated through "
    "photophosphorylation, and how the Calvin cycle uses these energy "
    "carriers to fix carbon dioxide into glucose. Include the role of stomata "
    "in gas exchange and discuss how environmental factors like light "
    "intensity, water availability, soil nutrients, atmospheric carbon "
    "dioxide concentration, and temperature affect the overall rate of "
    "photosynthesis. Discuss the differences between C3, C4, and CAM "
    "photosynthetic pathways and their evolutionary adaptations to different "
    "environmental conditions including drought stress and high light "
    "intensity. Finally explain how mitochondrial respiration interacts with "
    "chloroplast metabolism in the cells of plants and how this energy "
    "balance is maintained throughout the day and night cycle of plant "
    "tissues across different seasons in temperate and tropical biomes."
)
_NEAR_DUP_VARIANTS = [
    _NEAR_DUP_BASE,
    _NEAR_DUP_BASE.replace("process", "mechanism", 1),  # 1-word edit near start
    _NEAR_DUP_BASE.replace("glucose", "sugar", 1),       # 1-word edit in middle
]

# ── Class D: distinct singletons ────────────────────────────────────────────
_SINGLETONS = [
    "Compute the integral of sin(x) cos(x) from 0 to pi/2 using substitution.",
    "Describe the structure of a eukaryotic cell and its main organelles in detail.",
    "What is the wavelength of light with frequency 5 x 10^14 Hz in vacuum?",
    "Find the eigenvalues of the 2x2 matrix [[2, 1], [1, 2]] step by step.",
    "Explain how DNA replication occurs during cell division and list the enzymes involved.",
    "Balance the chemical equation H2 + O2 produces H2O and explain conservation of atoms.",
    "Determine the molarity of a solution containing 5 grams of NaCl in 250 mL of water.",
]


REALISTIC_SCIENCE_PROMPTS: list[str] = (
    [_EXACT_DUP, _EXACT_DUP, _EXACT_DUP]
    + _FORMATTING_VARIANTS
    + _NEAR_DUP_VARIANTS
    + _SINGLETONS
)

# Index ranges for assertions in tests.
EXACT_DUP_RANGE = range(0, 3)        # indices 0, 1, 2
FORMATTING_RANGE = range(3, 6)       # indices 3, 4, 5
NEAR_DUP_RANGE = range(6, 9)         # indices 6, 7, 8
SINGLETON_RANGE = range(9, 16)       # indices 9..15