評估 LLM - MLflow Evals、Auto Eval

搭配 MLflow 使用 LiteLLM

MLflow 提供一個 API mlflow.evaluate()，協助您評估您的 LLM https://mlflow.org/docs/latest/llms/llm-evaluate/index.html

前置需求

uv add litellm

uv add mlflow

步驟 1：在 CLI 上啟動 LiteLLM Proxy

LiteLLM 可讓您為所有支援的 LLM 建立 OpenAI 相容的伺服器。關於 litellm proxy 的更多資訊請見此處

$ litellm --model huggingface/bigcode/starcoder

#INFO: Proxy running on http://0.0.0.0:8000

以下是您如何為其他支援的 llm 建立 proxy

Bedrock

$ export AWS_ACCESS_KEY_ID=""
$ export AWS_REGION_NAME="" # e.g. us-west-2
$ export AWS_SECRET_ACCESS_KEY=""

$ litellm --model bedrock/anthropic.claude-v2

Huggingface (TGI)

$ export HUGGINGFACE_API_KEY=my-api-key #[OPTIONAL]

$ litellm --model huggingface/<your model name> --api_base https://k58ory32yinf1ly0.us-east-1.aws.endpoints.huggingface.cloud

Anthropic

$ export ANTHROPIC_API_KEY=my-api-key

$ litellm --model claude-instant-1

VLLM

假設您是在本機執行 vllm

$ litellm --model vllm/facebook/opt-125m

OpenAI Compatible Server

$ litellm --model openai/<model_name> --api_base <your-api-base>

TogetherAI

$ export TOGETHERAI_API_KEY=my-api-key

$ litellm --model together_ai/lmsys/vicuna-13b-v1.5-16k

Replicate

$ export REPLICATE_API_KEY=my-api-key

$ litellm \
  --model replicate/meta/llama-2-70b-chat:02e509c789964a7ea8736978a43525956ef40397be9033abf9fd2badfe68c9e3

Petals

$ litellm --model petals/meta-llama/Llama-2-70b-chat-hf

Palm

$ export PALM_API_KEY=my-palm-key

$ litellm --model palm/chat-bison

Azure OpenAI

$ export AZURE_API_KEY=my-api-key
$ export AZURE_API_BASE=my-api-base

$ litellm --model azure/my-deployment-name

AI21

$ export AI21_API_KEY=my-api-key

$ litellm --model j2-light

Cohere

$ export COHERE_API_KEY=my-api-key

$ litellm --model command-nightly

步驟 2：執行 MLflow

在執行評估之前，我們會將 openai.api_base 設定為步驟 1 的 litellm proxy

openai.api_base = "http://0.0.0.0:8000"

import openai
import pandas as pd
openai.api_key = "anything"             # this can be anything, we set the key on the proxy
openai.api_base = "http://0.0.0.0:8000" # set api base to the proxy from step 1


import mlflow
eval_data = pd.DataFrame(
    {
        "inputs": [
            "What is the largest country",
            "What is the weather in sf?",
        ],
        "ground_truth": [
            "India is a large country",
            "It's cold in SF today"
        ],
    }
)

with mlflow.start_run() as run:
    system_prompt = "Answer the following question in two sentences"
    logged_model_info = mlflow.openai.log_model(
        model="gpt-3.5",
        task=openai.ChatCompletion,
        artifact_path="model",
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": "{question}"},
        ],
    )

    # Use predefined question-answering metrics to evaluate our model.
    results = mlflow.evaluate(
        logged_model_info.model_uri,
        eval_data,
        targets="ground_truth",
        model_type="question-answering",
    )
    print(f"See aggregated evaluation results below: \n{results.metrics}")

    # Evaluation result for each data record is available in `results.tables`.
    eval_table = results.tables["eval_results_table"]
    print(f"See evaluation table below: \n{eval_table}")

MLflow 輸出

{'toxicity/v1/mean': 0.00014476531214313582, 'toxicity/v1/variance': 2.5759661361262862e-12, 'toxicity/v1/p90': 0.00014604929747292773, 'toxicity/v1/ratio': 0.0, 'exact_match/v1': 0.0}
Downloading artifacts: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1890.18it/s]
See evaluation table below:
                        inputs              ground_truth                                            outputs  token_count  toxicity/v1/score
0  What is the largest country  India is a large country   Russia is the largest country in the world in...           14           0.000146
1   What is the weather in sf?     It's cold in SF today   I'm sorry, I cannot provide the current weath...           36           0.000143

搭配 AutoEval 使用 LiteLLM

AutoEvals 是一個使用最佳實務快速且輕鬆評估 AI 模型輸出的工具。 https://github.com/braintrustdata/autoevals

前置需求

uv add litellm

uv add autoevals

快速開始

在這段程式碼範例中，我們使用來自 autoevals.llm 的 Factuality() 評估器，來測試輸出相較於原始（預期）值是否屬實。

Autoevals 預設使用 gpt-3.5-turbo / gpt-4-turbo 來評估回應

請參閱 autoevals 文件中關於支援的評估器的說明——Translation、Summary、Security Evaluators 等

# auto evals imports 
from autoevals.llm import *
###################
import litellm

# litellm completion call
question = "which country has the highest population"
response = litellm.completion(
    model = "gpt-3.5-turbo",
    messages = [
        {
            "role": "user",
            "content": question
        }
    ],
)
print(response)
# use the auto eval Factuality() evaluator
evaluator = Factuality()
result = evaluator(
    output=response.choices[0]["message"]["content"],       # response from litellm.completion()
    expected="India",                                       # expected output
    input=question                                          # question passed to litellm.completion
)

print(result)

評估輸出 - 來自 AutoEvals

Score(
    name='Factuality', 
    score=0, 
    metadata=
        {'rationale': "The expert answer is 'India'.\nThe submitted answer is 'As of 2021, China has the highest population in the world with an estimated 1.4 billion people.'\nThe submitted answer mentions China as the country with the highest population, while the expert answer mentions India.\nThere is a disagreement between the submitted answer and the expert answer.", 
        'choice': 'D'
        }, 
    error=None
)