An API for using metric models (either provided by default or fine-tuned yourself) to evaluate LLMs.
Project description
A library for using models (either default ones provided by LastMile or your own that are fine-tuned) to evaluate LLMs.
Evaluations are run on dataframes that include any combination of input
, ground_truth
, and output
columns. At least one of these columns must be defined and all values must be strings.
Example usage:
from lastmile_auto_eval import (
EvaluationMetric,
EvaluationResult,
evaluate,
stream_evaluate,
)
import pandas as pd
import json
from typing import Any, Generator
queries = ["what color is the sky?", "what color is the sky?"]
correct_answer = "the sky is blue"
incorrect_answer = "the sky is red"
ground_truth_values = [correct_answer, correct_answer]
responses = [correct_answer, incorrect_answer]
df = pd.DataFrame(
{
"input": queries,
"ground_truth": ground_truth_values,
"output": responses,
}
)
# Non-streaming
result: EvaluationResult = evaluate(
dataframe=df,
metrics=[
EvaluationMetric.P_FAITHFUL,
EvaluationMetric.SUMMARIZATION,
],
)
print(json.dumps(result, indent=2))
# Response will look something like this:
"""
{
"p_faithful": [
0.999255359172821,
0.00011296303273411468
],
"summarization": [
0.9995583891868591,
6.86283819959499e-05
]
}
"""
# Response-streaming
result_iterator: Generator[EvaluationResult, Any, Any] = (
stream_evaluate(
dataframe=df,
metrics=[
EvaluationMetric.P_FAITHFUL,
EvaluationMetric.SUMMARIZATION,
],
)
)
for result_chunk in result_iterator:
print(json.dumps(result_chunk, indent=2))
# Bidirectional-streaming
def gen_df_stream(input: list[str], gt: list[str], output: list[str]):
for i in range(len(input)):
df_chunk = pd.DataFrame(
{
"input": [input[i]],
"ground_truth": [gt[i]],
"output": [output[i]],
}
)
yield df_chunk
df_iterator = gen_df_stream(
input=queries, gt=ground_truth_values, output=responses
)
result_iterator: Generator[EvaluationResult, Any, Any] = (
stream_evaluate(
dataframe=df_iterator,
metrics=[
EvaluationMetric.P_FAITHFUL,
EvaluationMetric.SUMMARIZATION,
],
)
)
for result_chunk in result_iterator:
print(json.dumps(result_chunk, indent=2))
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Close
Hashes for lastmile_auto_eval-0.0.4-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 6321cb17238e06af9ad9fe6947958c6742430db0637142ef6c4a6dedc5604516 |
|
MD5 | 93bf10cd1b4cd699d2d5697376513581 |
|
BLAKE2b-256 | 3ec1d42e2cc81846d144b8a8e88033ff8c15aa6a3371ca354e69d00bd6d63114 |