Ragas Ragas
stable · 中文译文
中文译文 · 原文:https://docs.ragas.io/en/stable/howtos/cli/benchmark_llm/ · 许可证 Apache-2.0

LLM 基准测试 Quickstart

benchmark_llm 模板在折扣计算任务上对多种 LLM 模型做基准测试与比较。

创建项目

ragas quickstart benchmark_llm
cd benchmark_llm

安装依赖

uv sync

设置 API Key

export OPENAI_API_KEY="your-openai-key"
# Or other provider keys as needed

运行评测

uv run python evals.py

要基准测试特定模型:

uv run python evals.py --model gpt-4o
uv run python evals.py --model gpt-3.5-turbo

项目结构

benchmark_llm/
├── README.md              # Project documentation
├── pyproject.toml         # Project configuration
├── prompt.py              # Prompt implementation
├── evals.py               # Evaluation workflow
├── __init__.py            # Python package marker
└── evals/
    ├── datasets/
    │   └── discount_benchmark.csv  # Customer profiles and expected discounts
    ├── experiments/       # Evaluation results
    └── logs/              # Execution logs

评测内容

该模板在结构化输出任务上基准测试 LLM 表现:

  • 任务:根据客户画像计算折扣百分比
  • 模型:比较 GPT-4、GPT-3.5、Claude、Gemini 等
  • 输出格式:带折扣百分比的 JSON
  • 指标:折扣准确率(correct/incorrect)

理解代码

Prompt(prompt.py)

根据客户画像计算折扣:

from prompt import run_prompt

profile = "Premium customer, 5 years tenure, $50k annual spend"
result = await run_prompt(profile, model="gpt-4o")
# Returns: {"discount_percentage": 15}

评测(evals.py)

基准测试模型准确率:

@discrete_metric(name="discount_accuracy", allowed_values=["correct", "incorrect"])
def discount_accuracy(prediction: str, expected_discount):
    parsed_json = json.loads(prediction)
    predicted_discount = parsed_json.get("discount_percentage")

    if predicted_discount == int(expected_discount):
        return MetricResult(value="correct", ...)
    else:
        return MetricResult(value="incorrect", ...)

测试数据

模板包含 evals/datasets/discount_benchmark.csv,其中包括:

  • 客户画像(tenure、spend、tier 等)
  • 期望折扣百分比
  • 折扣计算的业务规则

基准测试多个模型

对相同评测在不同模型上运行:

# GPT-4
uv run python evals.py --model gpt-4o

# GPT-3.5
uv run python evals.py --model gpt-3.5-turbo

# Claude
uv run python evals.py --model claude-3-5-sonnet-20241022

# Compare results

定制

添加你自己的任务

修改 prompt 以基准测试不同能力:

# Code generation
prompt = "Generate Python code to {task}"

# Summarization
prompt = "Summarize this text in 50 words: {text}"

# Classification
prompt = "Classify this email as spam/not-spam: {email}"

比较成本与延迟

跟踪额外指标:

import time

start = time.time()
response = await run_prompt(profile, model=model_name)
latency = time.time() - start

# Log cost and latency alongside accuracy

分析结果

比较模型表现:

import pandas as pd

gpt4_results = pd.read_csv("evals/experiments/gpt4_benchmark.csv")
gpt35_results = pd.read_csv("evals/experiments/gpt35_benchmark.csv")

print(f"GPT-4 Accuracy: {(gpt4_results['discount_accuracy'] == 'correct').mean():.1%}")
print(f"GPT-3.5 Accuracy: {(gpt35_results['discount_accuracy'] == 'correct').mean():.1%}")

下一步