judgeval 0.0.1__py3-none-any.whl

judgeval/judges/mixture_of_judges.py

@@ -0,0 +1,248 @@
+"""
+Implementation for Mixture of Judges model through Judgeval
+
+Enables client to use multiple models to generate responses and then aggregate them into a single response.
+"""
+from judgeval import *
+import pydantic
+from typing import List, Union, Mapping, Dict
+from judgeval.judges import judgevalJudge
+from judgeval.common.utils import get_completion_multiple_models, get_chat_completion, aget_completion_multiple_models, aget_chat_completion
+from judgeval.common.logger import debug, error
+
+def build_dynamic_mixture_prompt(
+        judge_responses: List[str], 
+        custom_system_prompt: str = None, 
+        custom_conversation_history: List[Mapping] = None
+    ) -> List[Mapping]:
+    """
+    Dynamically builds a prompt to mix judge responses together for the Mixture of Judges model.
+
+    In this implementation, we simply concatenate the judge responses into a formatted string, then
+    pass it into a default prompt template. This template can be customized by providing a custom prompt.
+
+    Args:
+        judge_responses (List[str]): List of responses from individual judges to be synthesized
+        custom_system_prompt (str, optional): Custom system prompt to override the default one. Defaults to None.
+        custom_conversation_history (List[Mapping], optional): Custom conversation history to override the default one. Defaults to None.
+    """
+    formatted_responses = "\n".join([f"# Judge {i + 1}'s response: #\n{response}" for i, response in enumerate(judge_responses)])
+    
+    # This is the default prompt for the Mixture of Judges model
+    """
+    You are tasked with synthesizing responses from multiple expert judges. You will receive N individual answers on the same topic. Your job is to:
+
+    1. Analyze and compare the key points, patterns, and agreements between the answers.
+    2. Identify the consensus by focusing on areas where most or all of the answers align. Consider common reasoning and frequently mentioned conclusions.
+    3. Condense the responses into a single, coherent, and concise answer that represents the collective judgment of the group.
+    4. When opinions differ or contradict, highlight the most supported viewpoint while briefly acknowledging the dissenting perspectives.
+    5. Ensure the final answer is balanced and clear, providing a comprehensive summary that captures the wisdom of all judges while avoiding repetition.
+
+    ## Start of Judge Responses ##
+    {{judge_responses}}
+    ## End of Judge Responses ##
+    Synthesized response:
+    """
+
+    default_conversation = [  # inject the judge responses into the default prompt
+        {
+            'role': 'system',
+            'content': 'You are tasked with synthesizing responses from multiple expert judges. You will receive N individual answers on the same topic. Your job is to:\n1. Analyze and compare the key points, patterns, and agreements between the answers.\n2. Identify the consensus by focusing on areas where most or all of the answers align. Consider common reasoning and frequently mentioned conclusions.\n3. Condense the responses into a single, coherent, and concise answer that represents the collective judgment of the group.\n4. When opinions differ or contradict, highlight the most supported viewpoint while briefly acknowledging the dissenting perspectives.\n5. Ensure the final answer is balanced and clear, providing a comprehensive summary that captures the wisdom of all judges while avoiding repetition.\n\n**IMPORTANT**: IF THE JUDGE RESPONSES ARE IN JSON FORMAT, YOU MUST RESPOND USING THE SAME JSON FORMAT THAT THE RESPONSES ARE IN. If the judge responses are in JSON, you MUST RESPOND IN VALID JSON FORMAT. '
+        }, 
+        {
+            'role': 'user',
+            'content': '## Start of Judge Responses ## \n# Judge 1\'s response: #\n{\n"claims": [\n{\n"claim": "A 30-day full refund is offered.",\n"quote": "We offer a 30-day full refund at no extra cost."\n},\n{\n"claim": "The 30-day full refund comes at no extra cost.",\n"quote": "We offer a 30-day full refund at no extra cost."\n}\n]\n}\n\n# Judge 2\'s response: #\n{\n    "claims": [\n        {\n            "claim": "A full refund is offered within 30 days.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        },\n        {\n            "claim": "The 30-day full refund is offered at no extra cost.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        }\n    ]\n}\n# Judge 3\'s response: #\n {\n    "claims": [\n        {\n            "claim": "A 30-day full refund is offered.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        },\n        {\n            "claim": "The 30-day full refund is offered at no extra cost.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        }\n    ]\n}\n## End of Judge Responses ##\nSynthesized response:'
+        }, 
+        {
+            'role': 'assistant',
+            'content': 'The consensus among the judges is clear and unanimous. All three judges agree that a 30-day full refund is offered, and this refund is available at no extra cost. This conclusion is consistently supported by their statements, as each of their claims is directly quoted as: "We offer a 30-day full refund at no extra cost." There are no dissenting perspectives or opposing views provided in any of the responses, indicating complete alignment on this topic.\n\nJSON:\n{\n    "claims": [\n        {\n            "claim": "A full refund is offered within 30 days.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        },\n        {\n            "claim": "The 30-day full refund is offered at no extra cost.",\n            "quote": "We offer a 30-day full refund at no extra cost."\n        }\n    ]\n}'
+        },
+        {
+            'role': 'user',
+            'content': "## Start of Judge Responses ##\n# Judge 1's response: # \nThe capital of France is Paris.\n\n# Judge 2's response: #\nThe capital of France is Paris.\n\n# Judge 3's response: # \nThe capital of France is Paris. It's one of the most popular tourist destinations in the world, known for its art, culture, and history. It's also famous for its iconic landmarks such as the Eiffel Tower, Louvre Museum, and Notre-Dame Cathedral.\n\n## End of Judge Responses ##\nSynthesized response:"
+        },
+        {
+            'role': 'assistant',
+            'content': "The capital of France is Paris. It is widely recognized as one of the world's most popular tourist destinations, celebrated for its rich art, culture, and history. Paris is renowned for its iconic landmarks, including the Eiffel Tower, Louvre Museum, and Notre-Dame Cathedral."
+        }, 
+        {
+            'role': 'user',
+            'content': f'## Start of Judge Responses ##\n{formatted_responses}\n## End of Judge Responses ##\nSynthesized response:\n'
+        }
+    ]
+    
+    # If a custom system prompt is provided, validate and use it
+    if custom_system_prompt is not None:
+        if not isinstance(custom_system_prompt, str):
+            error(f"TypeError: Custom system prompt must be a string. Received: {type(custom_system_prompt)}.")
+            raise TypeError(f"Custom system prompt must be a string. Received: {type(custom_system_prompt)}.")
+        if not custom_system_prompt:
+            error("ValueError: Custom system prompt cannot be empty")
+            raise ValueError("Custom system prompt cannot be empty")
+        # Override the default system prompt, but also add special instructions for handling JSON
+        default_conversation[0]['content'] = custom_system_prompt + "\n\n**IMPORTANT**: IF THE JUDGE RESPONSES ARE IN JSON FORMAT, YOU MUST RESPOND USING THE SAME JSON FORMAT THAT THE RESPONSES ARE IN. If the judge responses are in JSON, you MUST RESPOND IN VALID JSON FORMAT."
+    
+    # If a custom conversation history is provided, append the judge responses to it
+    if custom_conversation_history is not None:
+        # Validate custom conversation history format
+        for message in custom_conversation_history:
+            if not isinstance(message, dict):
+                error(f"TypeError: Custom conversation history must be a list of dictionaries. Received: {message}.")
+                raise TypeError(f"Custom conversation history must be a list of dictionaries. Received: {message}.")
+            
+            if 'role' not in message or 'content' not in message:
+                error("ValueError: Each message must have 'role' and 'content' keys")
+                raise ValueError("Each message must have 'role' and 'content' keys")
+            
+            if not isinstance(message['role'], str) or not isinstance(message['content'], str):
+                error(f"TypeError: Message role and content must be strings. Received: {type(message['role'])}, {type(message['content'])}.")
+                raise TypeError(f"Message role and content must be strings. Received: {type(message['role'])}, {type(message['content'])}.")
+            
+            if message['role'] not in ['system', 'user', 'assistant']:
+                error(f"ValueError: Message role must be one of: 'system', 'user', 'assistant'. Received: {message['role']}.")
+                raise ValueError(f"Message role must be one of: 'system', 'user', 'assistant'. Received: {message['role']}.")
+            
+        judge_responses_prompt = {
+            'role': 'user', 
+            'content': f'## Start of Judge Responses ##\n{formatted_responses}\n## End of Judge Responses ##\nSynthesized response:\n'
+        }
+        return custom_conversation_history + [judge_responses_prompt]
+    
+    # Otherwise return the default conversation with system prompt and examples
+    # No customization, return the default conversation with system prompt and examples
+    return default_conversation
+
+BASE_CONVERSATION = [
+    {"role": "system", "content": "You are a helpful assistant."},
+]  # for string inputs, we need to add the user query to a base conversation, since LiteLLM only accepts a list of dictionaries as a chat history
+class MixtureOfJudges(judgevalJudge):
+    """
+    IMPORTANT: When supplying custom prompts and conversation histories for aggregation, supply them in the following format:
+    in kwargs:
+    {
+        "custom_prompt": "Your custom prompt here",
+        "custom_conversation": [
+            {"role": "system", "content": "System message 1"},
+            {"role": "user", "content": "User message 1"},
+            {"role": "assistant", "content": "Assistant message 1"},
+            ...
+        ]
+    }
+    """
+    def __init__(self, 
+                 models: List[str] = ['QWEN', 'LLAMA3_70B_INSTRUCT_TURBO', 'MISTRAL_8x22B_INSTRUCT'],
+                 aggregator: str = 'gpt-4o', 
+                 **kwargs):
+        """
+        `models` are the individual judge models to be used for generating responses.
+        `aggregator` is the model that will aggregate the responses from the individual judges.
+
+        kwargs include "custom_prompt" and "custom_conversation" for customizing the prompt for the Mixture of Judges model.
+        """
+        self.models = models
+        self.aggregator = aggregator
+        self.kwargs = kwargs
+        super().__init__(model_name=models)
+
+    def generate(
+            self, 
+            input: Union[str, List[Mapping[str, str]]], 
+            response_schema: pydantic.BaseModel = None, 
+            aggregation_schema: pydantic.BaseModel = None,
+            **kwargs) -> str:
+        """
+        Args:
+            input (Union[str, List[Mapping[str, str]]]): Input query or conversation history to the model.
+            response_schema (pydantic.BaseModel): Response schema for individual judge models.
+            aggregation_schema (pydantic.BaseModel): Response schema for the aggregator model.
+            kwargs: Additional keyword arguments.
+        """
+        debug(f"Generating response for input type: {type(input)}")
+        
+        # Convert input to conversation format if needed
+        if isinstance(input, str):
+            convo = BASE_CONVERSATION + [{"role": "user", "content": input}]
+        elif isinstance(input, list):
+            convo = input
+        else:
+            error(f"Invalid input type received: {type(input)}")
+            raise TypeError(f"Input must be a string or a list of dictionaries. Input type of: {type(input)}")
+
+        try:
+            responses = get_completion_multiple_models(
+                models=self.models,
+                messages=[convo] * len(self.models),
+                response_formats=[response_schema] * len(self.models)
+            )
+        except Exception as e:
+            error(f"Error getting completions from multiple models: {str(e)}")
+            raise
+
+        compiled_mixture_prompt = build_dynamic_mixture_prompt(responses, self.kwargs.get('custom_prompt'), self.kwargs.get('custom_conversation'))
+        
+        try:
+            mixed_response = get_chat_completion(
+                model_type=self.aggregator,
+                messages=compiled_mixture_prompt,
+                response_format=aggregation_schema,
+            )
+        except Exception as e:
+            error(f"Error getting chat completion from aggregator: {str(e)}")
+            raise
+            
+        return mixed_response
+
+    async def a_generate(
+            self, 
+            input: Union[str, List[Mapping[str, str]]], 
+            response_schema: pydantic.BaseModel = None,
+            aggregation_schema: pydantic.BaseModel = None,
+            **kwargs
+        ) -> str:
+        """
+        Args:
+            input (Union[str, List[Mapping[str, str]]]): Input query or conversation history to the model.
+            response_schema (pydantic.BaseModel): Response schema for individual judge models.
+            aggregation_schema (pydantic.BaseModel): Response schema for the aggregator model.
+            kwargs: Additional keyword arguments.
+        """
+        debug(f"Generating response for input type: {type(input)}")
+        
+        # Convert input to conversation format if needed
+        if isinstance(input, str):
+            convo = BASE_CONVERSATION + [{"role": "user", "content": input}]
+        elif isinstance(input, list):
+            convo = input
+        else:
+            error(f"Invalid input type received: {type(input)}")
+            raise TypeError(f"Input must be a string or a list of dictionaries. Input type of: {type(input)}")
+
+        try:
+            responses = await aget_completion_multiple_models(
+                models=self.models,
+                messages=[convo] * len(self.models),
+                response_formats=[response_schema] * len(self.models)
+            )
+        except Exception as e:
+            error(f"Error getting async completions from multiple models: {str(e)}")
+            raise
+
+        compiled_mixture_prompt = build_dynamic_mixture_prompt(responses, self.kwargs.get('custom_prompt'), self.kwargs.get('custom_conversation'))
+        
+        try:
+            mixed_response = await aget_chat_completion(
+                model_type=self.aggregator,
+                messages=compiled_mixture_prompt,
+                response_format=aggregation_schema,
+            )
+        except Exception as e:
+            error(f"Error getting async chat completion from aggregator: {str(e)}")
+            raise
+            
+        return mixed_response
+    
+    def load_model(self):
+        return self.models
+
+    def get_model_name(self) -> List[str]:
+        return self.models

judgeval/judges/together_judge.py

@@ -0,0 +1,55 @@
+"""
+Implementation of using TogetherAI inference for judges.
+"""
+
+from pydantic import BaseModel
+from typing import List, Union, Mapping
+from judgeval.common.logger import debug, error
+
+from judgeval.judges import judgevalJudge
+from judgeval.common.utils import fetch_together_api_response, afetch_together_api_response
+
+BASE_CONVERSATION = [
+    {"role": "system", "content": "You are a helpful assistant."},
+]
+
+class TogetherJudge(judgevalJudge):
+    def __init__(self, model: str = "QWEN", **kwargs):
+        debug(f"Initializing TogetherJudge with model={model}")
+        self.model = model
+        self.kwargs = kwargs
+        super().__init__(model_name=model)
+
+    # TODO: Fix cost for generate and a_generate
+    def generate(self, input: Union[str, List[Mapping[str, str]]], schema: BaseModel = None) -> str:
+        debug(f"Generating response for input type: {type(input)}")
+        if isinstance(input, str):
+            convo = BASE_CONVERSATION + [{"role": "user", "content": input}]
+            return fetch_together_api_response(self.model, convo, response_format=schema)
+        elif isinstance(input, list):
+            convo = input
+            return fetch_together_api_response(self.model, convo, response_format=schema)
+        else:
+            error(f"Invalid input type received: {type(input)}")
+            raise TypeError("Input must be a string or a list of dictionaries.")
+
+    async def a_generate(self, input: Union[str, List[dict]], schema: BaseModel = None) -> str:
+        debug(f"Async generating response for input type: {type(input)}")
+        if isinstance(input, str):
+            convo = BASE_CONVERSATION + [{"role": "user", "content": input}]
+            res = await afetch_together_api_response(self.model, convo, response_format=schema)
+            return res
+        elif isinstance(input, list):
+            convo = input
+            res = await afetch_together_api_response(self.model, convo, response_format=schema)
+            return res
+        else:
+            error(f"Invalid input type received: {type(input)}")
+            raise TypeError("Input must be a string or a list of dictionaries.")
+
+    def load_model(self) -> str:
+        return self.model
+
+    def get_model_name(self) -> str:
+        return self.model
+    

judgeval/judges/utils.py

@@ -0,0 +1,45 @@
+"""
+This module contains utility functions for judge models.
+"""
+import litellm
+from typing import Optional, Union, Tuple, List
+
+from judgeval.common.exceptions import InvalidJudgeModelError
+from judgeval.judges import judgevalJudge, LiteLLMJudge, TogetherJudge, MixtureOfJudges
+from judgeval.constants import TOGETHER_SUPPORTED_MODELS
+
+LITELLM_SUPPORTED_MODELS = set(litellm.model_list)
+
+def create_judge(
+    model: Optional[Union[str, List[str], judgevalJudge]] = None) -> Tuple[judgevalJudge, bool]:
+    """
+    Creates a judge model from string(s) or a judgeval judge object.
+
+    If `model` is a single string, it is assumed to be a judge model name.
+    If `model` is a list of strings, it is assumed to be a list of judge model names (for MixtureOfJudges).
+    If `model` is a judgeval judge object, it is returned as is.    
+
+    Returns a tuple of (initialized judgevalBaseLLM, using_native_model boolean)
+    If no model is provided, uses GPT4o as the default judge.
+    """
+    if model is None:  # default option
+        return LiteLLMJudge(model="gpt-4o"), True
+    if not isinstance(model, (str, list, judgevalJudge)):
+        raise InvalidJudgeModelError(f"Model must be a string, list of strings, or a judgeval judge object. Got: {type(model)} instead.")
+    # If model is already a valid judge type, return it and mark native
+    if isinstance(model, (judgevalJudge, LiteLLMJudge, TogetherJudge, MixtureOfJudges)):
+        return model, True 
+
+    # Either string or List[str]
+    if isinstance(model, list):
+        for m in model:
+            if m not in TOGETHER_SUPPORTED_MODELS and m not in LITELLM_SUPPORTED_MODELS:
+                raise InvalidJudgeModelError(f"Invalid judge model chosen: {m}")
+        return MixtureOfJudges(models=model), True
+    # If model is a string, check that it corresponds to a valid model
+    if model in LITELLM_SUPPORTED_MODELS:
+        return LiteLLMJudge(model=model), True
+    if model in TOGETHER_SUPPORTED_MODELS:
+        return TogetherJudge(model=model), True
+    else:
+        raise InvalidJudgeModelError(f"Invalid judge model chosen: {model}")

judgeval/judgment_client.py

@@ -0,0 +1,244 @@
+"""
+Implements the JudgmentClient to interact with the Judgment API.
+"""
+import os
+from typing import Optional, List, Dict, Any, Union
+import requests
+
+from judgeval.constants import ROOT_API
+from judgeval.data.datasets import EvalDataset
+from judgeval.data import ScoringResult, Example
+from judgeval.judges import judgevalJudge
+from judgeval.scorers import JudgmentScorer, CustomScorer, ClassifierScorer
+from judgeval.evaluation_run import EvaluationRun
+from judgeval.run_evaluation import run_eval
+from judgeval.constants import JUDGMENT_EVAL_FETCH_API_URL
+from judgeval.common.exceptions import JudgmentAPIError
+from pydantic import BaseModel
+
+class EvalRunRequestBody(BaseModel):
+    eval_name: str
+    project_name: str
+    judgment_api_key: str
+
+
+class JudgmentClient:
+    def __init__(self, judgment_api_key: str = os.getenv("JUDGMENT_API_KEY")):
+        self.judgment_api_key = judgment_api_key
+        
+        # Verify API key is valid
+        result, response = self._validate_api_key()
+        if not result:
+            # May be bad to output their invalid API key...
+            raise JudgmentAPIError(f"Issue with passed in Judgment API key: {response}")
+        else:
+            print(f"Successfully initialized JudgmentClient, welcome back {response.get('detail', {}).get('user_name', 'user')}!")
+            
+    def run_evaluation(
+        self, 
+        examples: List[Example],
+        scorers: List[Union[JudgmentScorer, CustomScorer]],
+        model: Union[str, List[str], judgevalJudge],
+        aggregator: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        log_results: bool = False,
+        project_name: str = "",
+        eval_run_name: str = "",
+        override: bool = False,
+    ) -> List[ScoringResult]:
+        """
+        Executes an evaluation of `Example`s using one or more `Scorer`s
+        """
+        try:
+            eval = EvaluationRun(
+                log_results=log_results,
+                project_name=project_name,
+                eval_name=eval_run_name,
+                examples=examples,
+                scorers=scorers,
+                model=model,
+                aggregator=aggregator,
+                metadata=metadata,
+                judgment_api_key=self.judgment_api_key
+            )
+            return run_eval(eval, override)
+        except ValueError as e:
+            raise ValueError(f"Please check your EvaluationRun object, one or more fields are invalid: \n{str(e)}")
+    
+    def evaluate_dataset(
+        self, 
+        dataset: EvalDataset,
+        scorers: List[Union[JudgmentScorer, CustomScorer]],
+        model: Union[str, List[str]],
+        aggregator: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        project_name: str = "",
+        eval_run_name: str = "",
+        log_results: bool = False
+    ) -> List[ScoringResult]:
+        """
+        Executes an evaluation of a `EvalDataset` using one or more `Scorer`s
+        """
+        try:
+            evaluation_run = EvaluationRun(
+                log_results=log_results,
+                project_name=project_name,
+                eval_name=eval_run_name,
+                examples=dataset.examples,
+                scorers=scorers,
+                model=model,
+                aggregator=aggregator,
+                metadata=metadata,
+                judgment_api_key=self.judgment_api_key
+            )
+            return run_eval(evaluation_run)
+        except ValueError as e:
+            raise ValueError(f"Please check your EvaluationRun object, one or more fields are invalid: \n{str(e)}")
+
+    def create_dataset(self) -> EvalDataset:
+        return EvalDataset(judgment_api_key=self.judgment_api_key)
+
+    def push_dataset(self, alias: str, dataset: EvalDataset, overwrite: Optional[bool] = False) -> bool:
+        """
+        Uploads an `EvalDataset` to the Judgment platform for storage.
+
+        Args:
+            alias (str): The name to use for the dataset
+            dataset (EvalDataset): The dataset to upload to Judgment
+            overwrite (Optional[bool]): Whether to overwrite the dataset if it already exists
+
+        Returns:
+            bool: Whether the dataset was successfully uploaded
+        """
+        # Set judgment_api_key just in case it was not set
+        dataset.judgment_api_key = self.judgment_api_key
+        return dataset.push(alias, overwrite)
+    
+    def pull_dataset(self, alias: str) -> EvalDataset:
+        """
+        Retrieves a saved `EvalDataset` from the Judgment platform.
+
+        Args:
+            alias (str): The name of the dataset to retrieve
+
+        Returns:
+            EvalDataset: The retrieved dataset
+        """
+        dataset = EvalDataset(judgment_api_key=self.judgment_api_key)
+        dataset.pull(alias)
+        return dataset
+    
+    # Maybe add option where you can pass in the EvaluationRun object and it will pull the eval results from the backend
+    def pull_eval(self, project_name: str, eval_run_name: str) -> List[Dict[str, Union[str, List[ScoringResult]]]]:
+        """Pull evaluation results from the server.
+
+        Args:
+            project_name (str): Name of the project
+            eval_run_name (str): Name of the evaluation run
+
+        Returns:
+            Dict[str, Union[str, List[ScoringResult]]]: Dictionary containing:
+                - id (str): The evaluation run ID
+                - results (List[ScoringResult]): List of scoring results
+        """
+        eval_run_request_body = EvalRunRequestBody(project_name=project_name, 
+                                                   eval_name=eval_run_name, 
+                                                   judgment_api_key=self.judgment_api_key)
+        eval_run = requests.post(JUDGMENT_EVAL_FETCH_API_URL,
+                                 json=eval_run_request_body.model_dump())
+        if eval_run.status_code != requests.codes.ok:
+            raise ValueError(f"Error fetching eval results: {eval_run.json()}")
+
+        eval_run_result = [{}]
+        for result in eval_run.json():
+            result_id = result.get("id", "")
+            result_data = result.get("result", dict())
+            filtered_result = {k: v for k, v in result_data.items() if k in ScoringResult.__annotations__}
+            eval_run_result[0]["id"] = result_id
+            eval_run_result[0]["results"] = [ScoringResult(**filtered_result)]
+        return eval_run_result
+        
+    def _validate_api_key(self):
+        """
+        Validates that the user api key is valid
+        """
+        response = requests.post(
+            f"{ROOT_API}/validate_api_key/",
+            json={"api_key": self.judgment_api_key}
+        )
+        if response.status_code == 200:
+            return True, response.json()
+        else:
+            return False, response.json().get("detail", "Error validating API key")
+
+    def fetch_classifier_scorer(self, slug: str) -> ClassifierScorer:
+        """
+        Fetches a classifier scorer configuration from the Judgment API.
+
+        Args:
+            slug (str): Slug identifier of the custom scorer to fetch
+
+        Returns:
+            ClassifierScorer: The configured classifier scorer object
+
+        Raises:
+            JudgmentAPIError: If the scorer cannot be fetched or doesn't exist
+        """
+        request_body = {
+            "slug": slug,
+            "judgment_api_key": self.judgment_api_key
+        }
+        
+        response = requests.post(
+            f"{ROOT_API}/fetch_scorer/",
+            json=request_body
+        )
+        
+        if response.status_code == 500:
+            raise JudgmentAPIError(f"The server is temporarily unavailable. Please try your request again in a few moments. Error details: {response.json().get('detail', '')}")
+        elif response.status_code != 200:
+            raise JudgmentAPIError(f"Failed to fetch classifier scorer '{slug}': {response.json().get('detail', '')}")
+            
+        scorer_config = response.json()
+        
+        try:
+            return ClassifierScorer(**scorer_config)
+        except Exception as e:
+            raise JudgmentAPIError(f"Failed to create classifier scorer '{slug}' with config {scorer_config}: {str(e)}")
+
+    def push_classifier_scorer(self, scorer: ClassifierScorer, slug: str = None) -> str:
+        """
+        Pushes a classifier scorer configuration to the Judgment API.
+
+        Args:
+            slug (str): Slug identifier for the scorer. If it exists, the scorer will be updated.
+            scorer (ClassifierScorer): The classifier scorer to save
+
+        Returns:
+            str: The slug identifier of the saved scorer
+
+        Raises:
+            JudgmentAPIError: If there's an error saving the scorer
+        """
+        request_body = {
+            "name": scorer.name,
+            "conversation": scorer.conversation,
+            "options": scorer.options,
+            "judgment_api_key": self.judgment_api_key,
+            "slug": slug
+        }
+        
+        response = requests.post(
+            f"{ROOT_API}/save_scorer/",
+            json=request_body
+        )
+        
+        if response.status_code == 500:
+            raise JudgmentAPIError(f"The server is temporarily unavailable. \
+                                   Please try your request again in a few moments. \
+                                   Error details: {response.json().get('detail', '')}")
+        elif response.status_code != 200:
+            raise JudgmentAPIError(f"Failed to save classifier scorer: {response.json().get('detail', '')}")
+            
+        return response.json()["slug"]
+