from __future__ import annotations
import regex as re
from langcheck.metrics.eval_clients import EvalClient
from langcheck.metrics.metric_inputs import (
get_metric_inputs,
get_metric_inputs_with_required_lists,
)
from langcheck.metrics.metric_value import MetricValue
from langcheck.metrics.scorer.hf_models import (
AutoModelForSequenceClassificationScorer,
)
from langcheck.utils.progress_bar import tqdm_wrapper
LANG = "ja"
[docs]
def sentiment(
generated_outputs: list[str] | str,
prompts: list[str] | str | None = None,
eval_model: str | EvalClient = "local",
local_overflow_strategy: str = "truncate",
) -> MetricValue[float | None]:
"""Calculates the sentiment scores of generated outputs. This metric takes
on float values between [0, 1], where 0 is negative sentiment and 1 is
positive sentiment. (NOTE: when using an EvalClient, the sentiment scores
are either 0.0 (negative), 0.5 (neutral), or 1.0 (positive). The score may
also be `None` if it could not be computed.)
We currently support two evaluation model types:
1. The 'local' type, where the Twitter-roBERTa-base-sentiment-multilingual
model is downloaded from HuggingFace and run locally. This is the default
model type and there is no setup needed to run this.
2. The EvalClient type, where you can use an EvalClient typically
implemented with an LLM. The implementation details are explained in each of
the concrete EvalClient classes.
Ref:
https://huggingface.co/cardiffnlp/twitter-xlm-roberta-base-sentiment-multilingual
Args:
generated_outputs: The model generated output(s) to evaluate
prompts: The prompts used to generate the output(s). Prompts are
optional metadata and not used to calculate the metric.
eval_model: The type of model to use ('local' or the EvalClient instance
used for the evaluation). default 'local'
local_overflow_strategy: The strategy to handle the inputs that are too
long for the local model. The supported strategies are 'nullify',
'truncate', and 'raise'. If 'nullify', the outputs that are too long
will be assigned a score of None. If 'truncate', the outputs that
are too long will be truncated. If 'raise', an error will be raised
when the outputs are too long. The default value is 'nullify'.
Returns:
An :class:`~langcheck.metrics.metric_value.MetricValue` object
"""
metric_inputs, [generated_outputs] = get_metric_inputs_with_required_lists(
generated_outputs=generated_outputs,
prompts=prompts,
required_params=["generated_outputs"],
)
metric_name = "sentiment"
if eval_model == "local":
scores = _sentiment_local(generated_outputs, local_overflow_strategy)
explanations = None
return MetricValue(
metric_name=metric_name,
metric_inputs=metric_inputs,
explanations=explanations,
metric_values=scores,
language=LANG,
)
else: # EvalClient
assert isinstance(
eval_model, EvalClient
), "An EvalClient must be provided for non-local model types."
sentiment_template = eval_model.load_prompt_template(
language=LANG, metric_name=metric_name
)
sentiment_assessment_to_score = {
"Positive": 1.0,
"Neutral": 0.5,
"Negative": 0.0,
}
return eval_model.compute_metric_values_from_template(
metric_inputs=metric_inputs,
template=sentiment_template,
metric_name=metric_name,
language=LANG,
score_map=sentiment_assessment_to_score,
)
def _sentiment_local(
generated_outputs: list[str], overflow_strategy: str
) -> list[float | None]:
"""Calculates the sentiment scores of generated outputs using the
Twitter-roBERTa-base-sentiment-multilingual model. This metric takes on
float values between [0, 1], where 0 is negative sentiment and 1 is positive
sentiment.
Ref:
https://huggingface.co/cardiffnlp/twitter-xlm-roberta-base-sentiment-multilingual
Args:
generated_outputs: A list of model generated outputs to evaluate
overflow_strategy: The strategy to handle inputs that are longer than
the maximum input length of the model.
Returns:
A list of scores
"""
scorer = AutoModelForSequenceClassificationScorer(
language="ja",
metric="sentiment",
# Each class represents a sentiment: 0 is negative, 1 is neutral, and 2
# is positive
class_weights=[0, 0.5, 1],
overflow_strategy=overflow_strategy,
max_input_length=512,
)
return scorer.score(generated_outputs)
[docs]
def toxicity(
generated_outputs: list[str] | str,
prompts: list[str] | str | None = None,
eval_model: str | EvalClient = "local",
local_overflow_strategy: str = "truncate",
eval_prompt_version: str = "v2",
) -> MetricValue[float | None]:
"""Calculates the toxicity scores of generated outputs. This metric takes on
float values between [0, 1], where 0 is low toxicity and 1 is high toxicity.
(NOTE: when using an EvalClient, the toxicity scores are either 0.0
(nontoxic), or 1.0 (toxic). The score may also be `None` if it could not be
computed.)
We currently support two evaluation model types:
1. The 'local' type, where a model file is downloaded from HuggingFace and
run locally. This is the default model type and there is no setup needed to
run this.
The model (Alnusjaponica/toxicity-score-multi-classification) is a
fine-tuned model based on line-corporation/line-distilbert-base-japanese
model.
2. The EvalClient type, where you can use an EvalClient typically
implemented with an LLM. The implementation details are explained in each of
the concrete EvalClient classes.
Ref:
https://huggingface.co/line-corporation/line-distilbert-base-japanese
https://huggingface.co/Alnusjaponica/toxicity-score-multi-classification
Args:
generated_outputs: The model generated output(s) to evaluate
prompts: The prompts used to generate the output(s). Prompts are
optional metadata and not used to calculate the metric.
eval_model: The type of model to use ('local' or the EvalClient instance
used for the evaluation). default 'local'
local_overflow_strategy: The strategy to handle the inputs that are too
long for the local model. The supported strategies are 'nullify',
'truncate', and 'raise'. If 'nullify', the outputs that are too long
will be assigned a score of None. If 'truncate', the outputs that
are too long will be truncated. If 'raise', an error will be raised
when the outputs are too long. The default value is 'nullify'.
eval_prompt_version: The version of the eval prompt to use when the
EvalClient is used. The default version is 'v2' (latest).
Returns:
An :class:`~langcheck.metrics.metric_value.MetricValue` object
"""
metric_inputs, [generated_outputs] = get_metric_inputs_with_required_lists(
generated_outputs=generated_outputs,
prompts=prompts,
required_params=["generated_outputs"],
)
metric_name = "toxicity"
if eval_model == "local":
scores = _toxicity_local(generated_outputs, local_overflow_strategy)
explanations = None
return MetricValue(
metric_name=metric_name,
metric_inputs=metric_inputs,
explanations=explanations,
metric_values=scores,
language=LANG,
)
else: # EvalClient
assert isinstance(
eval_model, EvalClient
), "An EvalClient must be provided for non-local model types."
toxicity_assessment_to_score = {
# The v1 prompt returns the toxicity on a scale of 1 to 5
"v1": {
"1": 0,
"2": 0.25,
"3": 0.5,
"4": 0.75,
"5": 1.0,
},
# The v2 prompt returns either "Toxic" or "Nontoxic"
"v2": {
"Toxic": 1.0,
"Nontoxic": 0,
},
}
assert (
eval_prompt_version in toxicity_assessment_to_score
), f"Invalid eval_prompt_version: {eval_prompt_version}. The valid versions are {list(toxicity_assessment_to_score.keys())}."
toxicity_template = eval_model.load_prompt_template(
language=LANG,
metric_name=metric_name,
eval_prompt_version=eval_prompt_version,
)
return eval_model.compute_metric_values_from_template(
metric_inputs=metric_inputs,
template=toxicity_template,
metric_name=metric_name,
language=LANG,
score_map=toxicity_assessment_to_score[eval_prompt_version],
)
def _toxicity_local(
generated_outputs: list[str], overflow_strategy: str
) -> list[float | None]:
"""Calculates the toxicity scores of generated outputs using a fine-tuned
model from `line-corporation/line-distilbert-base-japanese`. This metric
takes on float values between [0, 1], where 0 is low toxicity and 1 is high
toxicity.
Ref:
https://huggingface.co/line-corporation/line-distilbert-base-japanese
https://huggingface.co/Alnusjaponica/toxicity-score-multi-classification
Args:
generated_outputs: A list of model generated outputs to evaluate
overflow_strategy: The strategy to handle inputs that are longer than
the maximum input length of the model.
Returns:
A list of scores
"""
scorer = AutoModelForSequenceClassificationScorer(
language="ja",
metric="toxicity",
# The class 0 is for toxic texts.
class_weights=[1, 0],
overflow_strategy=overflow_strategy,
)
return scorer.score(generated_outputs)
[docs]
def fluency(
generated_outputs: list[str] | str,
prompts: list[str] | str | None = None,
eval_model: str | EvalClient = "local",
local_overflow_strategy: str = "truncate",
) -> MetricValue[float | None]:
"""Calculates the fluency scores of generated outputs. This metric takes on
float values between [0, 1], where 0 is low fluency and 1 is high fluency.
(NOTE: when using an EvalClient, the fluency scores are either 0.0
(poor), 0.5 (fair), or 1.0 (good). The score may also be `None` if it could
not be computed.)
We currently support two evaluation model types:
1. The 'local' type, where a model file is downloaded from HuggingFace and
run locally. This is the default model type and there is no setup needed to
run this.
The model (liwii/fluency-score-classification-ja) is a fine-tuned model
based on line-corporation/line-distilbert-base-japanese model.
2. The EvalClient type, where you can use an EvalClient typically
implemented with an LLM. The implementation details are explained in each of
the concrete EvalClient classes.
Ref:
https://huggingface.co/line-corporation/line-distilbert-base-japanese
https://huggingface.co/liwii/fluency-score-classification-ja
Args:
generated_outputs: The model generated output(s) to evaluate
prompts: The prompts used to generate the output(s). Prompts are
optional metadata and not used to calculate the metric.
eval_model: The type of model to use ('local' or the EvalClient instance
used for the evaluation). default 'local'
local_overflow_strategy: The strategy to handle the inputs that are too
long for the local model. The supported strategies are 'nullify',
'truncate', and 'raise'. If 'nullify', the outputs that are too long
will be assigned a score of None. If 'truncate', the outputs that
are too long will be truncated. If 'raise', an error will be raised
when the outputs are too long. The default value is 'nullify'.
Returns:
An :class:`~langcheck.metrics.metric_value.MetricValue` object
"""
metric_inputs, [generated_outputs] = get_metric_inputs_with_required_lists(
generated_outputs=generated_outputs,
prompts=prompts,
required_params=["generated_outputs"],
)
metric_name = "fluency"
if eval_model == "local":
scores = _fluency_local(generated_outputs, local_overflow_strategy)
explanations = None
return MetricValue(
metric_name=metric_name,
metric_inputs=metric_inputs,
explanations=explanations,
metric_values=scores,
language=LANG,
)
else: # EvalClient
assert isinstance(
eval_model, EvalClient
), "An EvalClient must be provided for non-local model types."
fluency_template = eval_model.load_prompt_template(
language=LANG, metric_name=metric_name
)
fluency_assessment_to_score = {
"Poor": 0,
"Fair": 0.5,
"Good": 1.0,
}
return eval_model.compute_metric_values_from_template(
metric_inputs=metric_inputs,
template=fluency_template,
metric_name=metric_name,
language=LANG,
score_map=fluency_assessment_to_score,
)
def _fluency_local(
generated_outputs: list[str], overflow_strategy: str
) -> list[float | None]:
"""Calculates the fluency scores of generated outputs using a fine-tuned
model from `line-corporation/line-distilbert-base-japanese`. This metric
takes on float values between [0, 1], where 0 is low fluency and 1 is high
fluency.
Ref:
https://huggingface.co/line-corporation/line-distilbert-base-japanese
https://huggingface.co/liwii/fluency-score-classification-ja
Args:
generated_outputs: A list of model generated outputs to evaluate
overflow_strategy: The strategy to handle inputs that are longer than
the maximum input length of the model.
Returns:
A list of scores
"""
scorer = AutoModelForSequenceClassificationScorer(
language="ja",
metric="fluency",
# The class 1 is for fluent texts.
class_weights=[0, 1],
overflow_strategy=overflow_strategy,
)
return scorer.score(generated_outputs)
[docs]
def tateishi_ono_yamada_reading_ease(
generated_outputs: list[str] | str,
prompts: list[str] | str | None = None,
) -> MetricValue[float]:
"""Calculates the readability of generated Japanese outputs using the
reading ease score introduced in "日本文の読みやすさの評価式 (A Computer
Readability Formula of Japanese Texts for Machine Scoring)". This metric
takes on float values between (-∞, ∞), but in the paper it is reported that
the average & the standard deviation of the scores obtained for 77 texts
used for the experiment are 50 and 10 respectively. Higher scores mean the
text is easier to read.
The score is based on the number of "run"s, which are sequences of
characters with the same type (hiragana, katakana, kanji... etc). See the
original paper for details.
Ref:
https://www.jstage.jst.go.jp/article/nihongokyoiku/158/0/158_49/_pdf/-char/ja (Japanese)
https://ipsj.ixsq.nii.ac.jp/ej/?action=pages_view_main&active_action=repository_view_main_item_detail&item_id=37773&item_no=1&page_id=13&block_id=8 (Japanese)
https://aclanthology.org/C88-2135/ (English)
Args:
generated_outputs: The model generated output(s) to evaluate
prompts: The prompts used to generate the output(s). Prompts are
optional metadata and not used to calculate the metric.
Returns:
An :class:`~langcheck.metrics.metric_value.MetricValue` object
"""
metric_inputs, [generated_outputs] = get_metric_inputs_with_required_lists(
generated_outputs=generated_outputs,
prompts=prompts,
required_params=["generated_outputs"],
)
# Regular expressions used to compute the reading ease score
blank_re = r"[ | |\n]"
hiragana_run_re = r"[\u3041-\u309F]+"
katakana_run_re = r"[\u30A1-\u30FE]+"
alphanumeric_run_re = r"[a-zA-Zａ-ｚＡ-Ｚ0-9０-９]+"
kanji_run_re = r"[\u4E00-\u9FFF]+"
delimiters_re = r"[、|。|!|？|!|?|「|」|,|，|.|．|…|『|』]"
# Aux function to compute the average length of strings in the list
def _mean_str_length(ls: list[str]) -> float:
if len(ls) == 0:
return 0
lens = [len(el) for el in ls]
return sum(lens) / len(lens)
def _get_reading_ease(text: str) -> float:
"""Computes reading ease for each example"""
# Preprocess the text: Delete all blanks
text = re.sub(blank_re, "", text)
# Get each term
hiragana_runs = re.findall(hiragana_run_re, text)
katakana_runs = re.findall(katakana_run_re, text)
alphanumeric_runs = re.findall(alphanumeric_run_re, text)
kanji_runs = re.findall(kanji_run_re, text)
sentences = re.split(delimiters_re, text)
period_count = text.count("。")
if period_count == 0:
# Just ignore the term
comma_period_ratio = 0
else:
comma_period_ratio = text.count("、") / period_count
return (
-0.12 * _mean_str_length(sentences)
- 1.37 * _mean_str_length(alphanumeric_runs)
+ 7.4 * _mean_str_length(hiragana_runs)
- 23.18 * _mean_str_length(kanji_runs)
- 5.3 * _mean_str_length(katakana_runs)
- 4.6 * comma_period_ratio
+ 115.79
)
scores = [
_get_reading_ease(text) for text in tqdm_wrapper(generated_outputs)
]
return MetricValue(
metric_name="tateishi_ono_yamada_reading_ease",
metric_inputs=metric_inputs,
explanations=None,
metric_values=scores,
language="ja",
)
[docs]
def jailbreak_prompt(
prompts: list[str] | str,
eval_model: EvalClient,
) -> MetricValue[float | None]:
"""Calculates whether jailbreak techniques are included in the prompts.
This metric takes on float values of either 0.0 (Low Risk),
0.5 (Medium Risk), or 1.0 (High Risk). The score may also be `None`
if it could not be computed.
We currently only support the evaluation based on an EvalClient.
"""
metric_inputs = get_metric_inputs(
prompts=prompts,
required_params=["prompts"],
)
metric_name = "jailbreak_prompt"
jailbreak_prompt_template = eval_model.load_prompt_template(
language=LANG, metric_name=metric_name
)
return eval_model.compute_metric_values_from_template(
metric_inputs=metric_inputs,
template=jailbreak_prompt_template,
metric_name=metric_name,
language=LANG,
score_map={
"Low Risk": 0.0,
"Medium Risk": 0.5,
"High Risk": 1.0,
},
)
[docs]
def prompt_leakage(
generated_outputs: list[str] | str,
system_prompts: list[str] | str,
eval_model: EvalClient,
eval_prompt_version: str = "v2",
) -> MetricValue[float | None]:
"""Calculates the severity of prompt leakage in the generated outputs.
This metric takes on float values of either 0.0 (Low Risk),
0.5 (Medium Risk), or 1.0 (High Risk). The score may also be `None`
if it could not be computed.
We currently only support the evaluation based on an EvalClient.
"""
if eval_prompt_version not in ["v1", "v2"]:
raise ValueError(
f"Invalid eval_prompt_version: {eval_prompt_version}. The valid versions are 'v1' and 'v2'."
)
metric_inputs = get_metric_inputs(
generated_outputs=generated_outputs,
additional_inputs={
"system_prompts": system_prompts,
},
additional_input_name_to_prompt_var_mapping={
"system_prompts": "system_prompt",
},
required_params=["generated_outputs", "system_prompts"],
)
metric_name = "prompt_leakage"
prompt_leakage_template = eval_model.load_prompt_template(
language=LANG,
metric_name=metric_name,
eval_prompt_version=eval_prompt_version,
)
return eval_model.compute_metric_values_from_template(
metric_inputs=metric_inputs,
template=prompt_leakage_template,
metric_name=metric_name,
language=LANG,
score_map={
"Low Risk": 0.0,
"Medium Risk": 0.5,
"High Risk": 1.0,
},
)
