What if I do not have an LLM service to run LLM-as-judge metrics?
No problem, just don’t include an eval_llm_client or an eval_embedding_client argument in the call(s) to the evaluation helpers. The helpers will automatically exclude any metrics that depend on them.
# BEFORE : default text metrics including those requiring target (ground_truth) and LLM-as-judge
text_metrics = dbnl.eval.metrics.text_metrics(
prediction="prediction", target="ground_truth", eval_llm_client=oai_client
)
# AFTER : remove the eval_llm_client to exclude LLM-as-judge metrics
text_metrics = dbnl.eval.metrics.text_metrics(
prediction="prediction", target="ground_truth"
)
aug_eval_df = evaluate(eval_df, text_metrics)
What if I do not have ground-truth available?
No problem. You can simply remove the target argument from the helper. The metric set helper will automatically exclude any metrics that depend on the target column being specified.
# BEFORE : default text metrics, including those requiring target (ground_truth) and LLM-as-judge
text_metrics = dbnl.eval.metrics.text_metrics(
prediction="prediction", target="ground_truth", eval_llm_client=oai_client
)
# AFTER : remove the target to remove metrics that depend on that value being specified
text_metrics = dbnl.eval.metrics.text_metrics(
prediction="prediction", eval_llm_client=oai_client
)
aug_eval_df = evaluate(eval_df, text_metrics)
There is an additional helper that can generate a list of generic metrics appropriate for “monitoring” unstructured text columns : text_monitor_metrics(). Simply provide a list of text column names and optionally an eval_llm_client for LLM-as-judge metrics.
# get text metrics for each column in list
monitor_metrics = dbnl.eval.metrics.text_monitor_metrics(
["prediction", "input"], eval_llm_client=oai_client
)
aug_eval_df = evaluate(eval_df, monitor_metrics)
How do I create a custom LLM-as-judge metric?
You can write your own LLM-as-judge metric that uses your custom prompt. The example below defines a custom LLM-as-judge metric and runs it on an example dataframe.
import dbnl
import os
import pandas as pd
from openai import OpenAI
from dbnl.eval.llm import OpenAILLMClient
from dbnl.eval import evaluate
from dbnl.eval.metrics.mlflow import MLFlowGenAIFromPromptEvaluationMetric
from dbnl.eval.metrics.metric import Metric
from dbnl.eval.llm.client import LLMClient
# 1. create client to power LLM-as-judge metrics
base_oai_client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
oai_client = OpenAILLMClient.from_existing_client(base_oai_client, llm_model="gpt-3.5-turbo-0125")
eval_df = pd.DataFrame(
[
{ "prediction":"France has no capital",
"ground_truth": "The capital of France is Paris",},
{ "prediction":"The capital of France is Toronto",
"ground_truth": "The capital of France is Paris",},
{ "prediction":"Paris is the capital",
"ground_truth": "The capital of France is Paris",},
] * 4
)
# 2. define a custom LLM-as-judge metric
def custom_text_similarity(prediction: str, target: str, eval_llm_client: LLMClient) -> Metric:
custom_prompt_v0 = """
Given the generated text : {prediction}, score the semantic similarity to the reference text : {target}.
Rate the semantic similarity from 1 (completely different meaning and facts between the generated and reference texts) to 5 (nearly the exact same semantic meaning and facts present in the generated and reference texts).
Example output, make certain that 'score:' and 'justification:' text is present in output:
score: 4
justification: XYZ
"""
return MLFlowGenAIFromPromptEvaluationMetric(
name="custom_text_similarity",
judge_prompt=custom_prompt_v0,
prediction=prediction,
target=target,
eval_llm_client=eval_llm_client,
version="v0",
)
# 3. instantiate the custom LLM-as-judge metric
c_metric = custom_text_similarity(
prediction='prediction', target='ground_truth', eval_llm_client=oai_client
)
# 4. run only the custom LLM-as-judge metric
aug_eval_df = evaluate(eval_df, [c_metric])
You can also write a metric that includes only the prediction column specified and reference only {prediction} in the custom prompt. An example is below:
def custom_text_simplicity(prediction: str, target: str, eval_llm_client: LLMClient) -> Metric:
custom_prompt_v0 = """
Given the generated text : {prediction}, score the text from 1 to 5 based on whether it is written in simple, easy to understand english
Rate the generated text from 5 (completely simple english, very commonly used words, easy to explain vocabulary) to 1 (complex english, uncommon words, difficult to explain vocabulary).
Example output, make certain that 'score:' and 'justification:' text is present in output:
score: 4
justification: XYZ
"""
return MLFlowGenAIFromPromptEvaluationMetric(
name="custom_text_simplicity",
judge_prompt=custom_prompt_v0,
prediction=prediction,
target=target,
eval_llm_client=eval_llm_client,
version="v0",
)
