File size: 4,641 Bytes
91bd31b
 
4d97278
91bd31b
 
 
 
 
 
 
4d97278
 
91bd31b
4d97278
 
 
 
 
91bd31b
 
 
 
 
 
 
4d97278
91bd31b
 
 
 
 
 
4d97278
91bd31b
 
 
 
 
 
4d97278
 
91bd31b
 
 
 
 
 
4d97278
91bd31b
 
 
 
 
 
4d97278
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91bd31b
 
 
4d97278
91bd31b
 
 
 
 
 
 
 
 
4d97278
91bd31b
 
 
 
 
 
 
4d97278
91bd31b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d97278
 
 
 
 
 
 
 
91bd31b
 
 
 
 
 
4d97278
 
 
 
 
 
 
 
 
 
91bd31b
 
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
import random
from dataclasses import dataclass
from typing import List
from enum import Enum

TASK_TYPES = ["summarization", "qa", "coding", "translation", "classification"]
INPUT_LENGTHS = ["short", "medium", "long"]


class DriftType(str, Enum):
    PROMPT_TEMPLATE_CHANGE = "prompt_template_change"
    QUANTIZATION_APPLIED = "quantization_applied"
    SAFETY_FILTER_MISCONFIG = "safety_filter_misconfig"
    DATA_CONTAMINATION = "data_contamination"
    CONTEXT_WINDOW_BUG = "context_window_bug"
    INFRA_LATENCY = "infra_latency"
    ROUTER_BUG = "router_bug"


DRIFT_CATALOG = {
    DriftType.PROMPT_TEMPLATE_CHANGE: {
        "description": "Upstream prompt template was silently modified",
        "affected_task_types": ["all"],
        "affected_input_lengths": ["all"],
        "quality_delta": -0.18,
        "remediation": "rollback_prompt_template",
    },
    DriftType.QUANTIZATION_APPLIED: {
        "description": "4-bit quantization applied to reduce serving cost",
        "affected_task_types": ["coding", "qa"],
        "affected_input_lengths": ["long"],
        "quality_delta": -0.25,
        "remediation": "revert_quantization",
    },
    DriftType.SAFETY_FILTER_MISCONFIG: {
        "description": "Safety filter misconfigured — false positives on sensitive keywords",
        "affected_task_types": ["all"],
        "affected_input_lengths": ["all"],
        "quality_delta": -0.12,
        "affected_probability": 0.3,
        "remediation": "recalibrate_safety_filter",
    },
    DriftType.DATA_CONTAMINATION: {
        "description": "Fine-tuning batch contaminated with low-quality summarization data",
        "affected_task_types": ["summarization"],
        "affected_input_lengths": ["all"],
        "quality_delta": -0.30,
        "remediation": "rollback_finetune_checkpoint",
    },
    DriftType.CONTEXT_WINDOW_BUG: {
        "description": "Context window handling bug truncates long inputs incorrectly",
        "affected_task_types": ["all"],
        "affected_input_lengths": ["long"],
        "quality_delta": -0.22,
        "remediation": "patch_context_window_handler",
    },
    DriftType.INFRA_LATENCY: {
        "description": "Serving infrastructure latency spike causing timeout-induced quality drops",
        "affected_task_types": ["coding", "summarization"],
        "affected_input_lengths": ["medium", "long"],
        "quality_delta": -0.15,
        "remediation": "scale_serving_infra",
    },
    DriftType.ROUTER_BUG: {
        "description": "Traffic router sending requests to a stale model checkpoint",
        "affected_task_types": ["all"],
        "affected_input_lengths": ["all"],
        "quality_delta": -0.20,
        "affected_probability": 0.4,
        "remediation": "fix_traffic_routing",
    },
}


@dataclass
class SimulatedOutput:
    sample_id: str
    task_type: str
    input_length: str
    base_quality: float
    final_quality: float
    drift_applied: List[str]


class DriftSimulator:
    def __init__(self, seed: int = 42):
        self.rng = random.Random(seed)

    def generate_pool(
        self,
        n: int = 200,
        active_drifts: List[DriftType] | None = None,
    ) -> List[SimulatedOutput]:
        pool = []
        active_drifts = active_drifts or []

        for i in range(n):
            task = self.rng.choice(TASK_TYPES)
            length = self.rng.choice(INPUT_LENGTHS)
            base_q = round(self.rng.gauss(0.78, 0.08), 3)
            base_q = max(0.0, min(1.0, base_q))

            final_q = base_q
            applied = []

            for drift in active_drifts:
                spec = DRIFT_CATALOG[drift]
                task_match = (
                    spec["affected_task_types"] == ["all"]
                    or task in spec["affected_task_types"]
                )
                len_match = (
                    spec.get("affected_input_lengths") == ["all"]
                    or length in spec.get("affected_input_lengths", ["all"])
                )
                prob_match = self.rng.random() < spec.get("affected_probability", 1.0)

                if task_match and len_match and prob_match:
                    final_q = max(0.0, final_q + spec["quality_delta"])
                    applied.append(drift.value)

            pool.append(
                SimulatedOutput(
                    sample_id=f"sample_{i:04d}",
                    task_type=task,
                    input_length=length,
                    base_quality=base_q,
                    final_quality=round(final_q, 3),
                    drift_applied=applied,
                )
            )

        return pool