如何定义自定义评估器

主要概念

• 评估器

自定义评估器是函数，它们接受数据集示例和生成的应用程序输出，并返回一个或多个指标。这些函数可以直接传递给 evaluate() / aevaluate()。

基本示例

Python

需要 langsmith>=0.2.0

from langsmith import evaluate

def correct(outputs: dict, reference_outputs: dict) -> bool:
"""Check if the answer exactly matches the expected answer."""
    return outputs["answer"] == reference_outputs["answer"]

def dummy_app(inputs: dict) -> dict:
    return {"answer": "hmm i'm not sure", "reasoning": "i didn't understand the question"}

results = evaluate(
    dummy_app,
    data="dataset_name",
    evaluators=[correct]
)

TypeScript

需要 langsmith>=0.2.9

import type { EvaluationResult } from "langsmith/evaluation";

const correct = async ({ outputs, referenceOutputs }: {
  outputs: Record<string, any>;
  referenceOutputs?: Record<string, any>;
}): Promise<EvaluationResult> => {
  const score = outputs?.answer === referenceOutputs?.answer;
  return { key: "correct", score };
}

评估器参数

自定义评估器函数必须具有特定的参数名称。它们可以接受以下参数的任意子集

• run: Run：应用程序根据给定示例生成的完整 Run 对象。
• example: Example：完整的数据集 Example，包括示例输入、输出（如果可用）和元数据（如果可用）。
• inputs: dict：数据集中与单个示例对应的输入字典。
• outputs: dict：应用程序根据给定 inputs 生成的输出字典。
• reference_outputs/referenceOutputs: dict：与示例关联的参考输出字典（如果可用）。

对于大多数用例，您只需要 inputs、outputs 和 reference_outputs。run 和 example 仅在您需要实际应用程序输入和输出之外的额外跟踪或示例元数据时才有用。

使用 JS/TS 时，这些都应作为单个对象参数的一部分传入。

评估器输出

自定义评估器预期返回以下类型之一

Python 和 JS/TS

• dict：形式为 {"score" | "value": ..., "key": ...} 的字典允许您自定义指标类型（“score”用于数值，“value”用于分类）和指标名称。例如，如果您想将整数记录为分类指标，这会很有用。

仅限 Python

• int | float | bool：这被解释为可以平均、排序等的连续指标。函数名称用作指标的名称。
• str：这被解释为分类指标。函数名称用作指标的名称。
• list[dict]：使用单个函数返回多个指标。

其他示例

Python

需要 langsmith>=0.2.0

from langsmith import evaluate, wrappers
from langsmith.schemas import Run, Example
from openai import AsyncOpenAI
# Assumes you've installed pydantic.
from pydantic import BaseModel

# We can still pass in Run and Example objects if we'd like
def correct_old_signature(run: Run, example: Example) -> dict:
    """Check if the answer exactly matches the expected answer."""
    return {"key": "correct", "score": run.outputs["answer"] == example.outputs["answer"]}

# Just evaluate actual outputs
def concision(outputs: dict) -> int:
    """Score how concise the answer is. 1 is the most concise, 5 is the least concise."""
    return min(len(outputs["answer"]) // 1000, 4) + 1

# Use an LLM-as-a-judge
oai_client = wrappers.wrap_openai(AsyncOpenAI())

async def valid_reasoning(inputs: dict, outputs: dict) -> bool:
    """Use an LLM to judge if the reasoning and the answer are consistent."""

    instructions = """\

Given the following question, answer, and reasoning, determine if the reasoning for the \
answer is logically valid and consistent with question and the answer."""

    class Response(BaseModel):
        reasoning_is_valid: bool

    msg = f"Question: {inputs['question']}\nAnswer: {outputs['answer']}\nReasoning: {outputs['reasoning']}"
    response = await oai_client.beta.chat.completions.parse(
        model="gpt-4o-mini",
        messages=[{"role": "system", "content": instructions,}, {"role": "user", "content": msg}],
        response_format=Response
    )
    return response.choices[0].message.parsed.reasoning_is_valid

def dummy_app(inputs: dict) -> dict:
    return {"answer": "hmm i'm not sure", "reasoning": "i didn't understand the question"}

results = evaluate(
    dummy_app,
    data="dataset_name",
    evaluators=[correct_old_signature, concision, valid_reasoning]
)

TypeScript

import { Client } from "langsmith";
import { evaluate } from "langsmith/evaluation";
import { Run, Example } from "langsmith/schemas";
import OpenAI from "openai";

// Type definitions
interface AppInputs {
    question: string;
}

interface AppOutputs {
    answer: string;
    reasoning: string;
}

interface Response {
    reasoning_is_valid: boolean;
}

// Old signature evaluator
function correctOldSignature(run: Run, example: Example) {
    return {
        key: "correct",
        score: run.outputs?.["answer"] === example.outputs?.["answer"],
    };
}

// Output-only evaluator
function concision({ outputs }: { outputs: AppOutputs }) {
    return {
        key: "concision",
        score: Math.min(Math.floor(outputs.answer.length / 1000), 4) + 1,
    };
}

// LLM-as-judge evaluator
const openai = new OpenAI();

async function validReasoning({
    inputs,
    outputs
}: {
    inputs: AppInputs;
    outputs: AppOutputs;
}) {
    const instructions = `      Given the following question, answer, and reasoning, determine if the reasoning for the       answer is logically valid and consistent with question and the answer.`;

    const msg = `Question: ${inputs.question}
Answer: ${outputs.answer}\nReasoning: ${outputs.reasoning}`;

    const response = await openai.chat.completions.create({
        model: "gpt-4",
        messages: [
            { role: "system", content: instructions },
            { role: "user", content: msg }
        ],
        response_format: { type: "json_object" },
        functions: [{
        name: "parse_response",
        parameters: {
            type: "object",
            properties: {
            reasoning_is_valid: {
                type: "boolean",
                description: "Whether the reasoning is valid"
            }
            },
            required: ["reasoning_is_valid"]
        }
        }]
    });

    const parsed = JSON.parse(response.choices[0].message.content ?? "{}") as Response;

    return {
        key: "valid_reasoning",
        score: parsed.reasoning_is_valid ? 1 : 0
    };
}

// Example application
function dummyApp(inputs: AppInputs): AppOutputs {
    return {
        answer: "hmm i'm not sure",
        reasoning: "i didn't understand the question"
    };
}

const results = await evaluate(dummyApp, {
        data: "dataset_name",
        evaluators: [correctOldSignature, concision, validReasoning],
        client: new Client()
});

相关内容

• 评估聚合实验结果：定义摘要评估器，用于计算整个实验的指标。
• 运行比较两个实验的评估：定义成对评估器，通过比较两个（或更多）实验来计算指标。