微调成本与收益分析

微调不是免费的——GPU 租赁、数据准备、迭代调优都需要投入。本章帮你量化成本，做出理性决策。

成本结构

GPU 训练成本参考

基于 1000 条训练样本，3 个 epoch。实际因数据量和超参数而异。

ROI 计算器

"""
微调 ROI 计算器
"""
from dataclasses import dataclass
@dataclass
class FinetuningCost:
"""微调成本模型"""
data_preparation_hours: float     # 数据准备人时
hourly_rate_usd: float = 50.0     # 人力时薪
gpu_cost_per_hour: float = 3.0    # GPU 单价
training_hours: float = 8.0       # 训练时长
num_iterations: int = 3           # 迭代次数
@property
def total_data_cost(self) -> float:
return self.data_preparation_hours * self.hourly_rate_usd
@property
def total_compute_cost(self) -> float:
return self.gpu_cost_per_hour * self.training_hours * self.num_iterations
@property
def total_cost(self) -> float:
return self.total_data_cost + self.total_compute_cost
@dataclass
class InferenceSavings:
"""推理节省模型"""
daily_requests: int
current_cost_per_request: float    # 使用大模型
finetuned_cost_per_request: float  # 使用微调小模型
@property
def daily_savings(self) -> float:
return self.daily_requests * (
self.current_cost_per_request - self.finetuned_cost_per_request
)
@property
def monthly_savings(self) -> float:
return self.daily_savings * 30
def calculate_roi(cost: FinetuningCost, savings: InferenceSavings) -> dict:
"""计算微调 ROI"""
total_investment = cost.total_cost
monthly_savings = savings.monthly_savings
payback_months = (
total_investment / monthly_savings if monthly_savings > 0 else float("inf")
)
annual_roi = (
(monthly_savings * 12 - total_investment) / total_investment * 100
)
return {
"total_investment_usd": round(total_investment, 2),
"monthly_savings_usd": round(monthly_savings, 2),
"payback_months": round(payback_months, 1),
"annual_roi_pct": round(annual_roi, 1),
"recommendation": (
"强烈推荐" if payback_months < 2
else "推荐" if payback_months < 6
else "需谨慎评估" if payback_months < 12
else "不建议"
),
}
# 示例：客服场景
cost = FinetuningCost(
data_preparation_hours=40,
training_hours=4,
num_iterations=3,
gpu_cost_per_hour=3.0,
)
savings = InferenceSavings(
daily_requests=10000,
current_cost_per_request=0.01,     # GPT-4
finetuned_cost_per_request=0.001,  # 微调 7B 模型
)
roi = calculate_roi(cost, savings)
# {'total_investment_usd': 2036.0, 'monthly_savings_usd': 2700.0,
#  'payback_months': 0.8, 'annual_roi_pct': 1492.4, 'recommendation': '强烈推荐'}

决策矩阵

本章小结

下一章：数据收集与格式化