Judges

TRL Judges is an experimental API which is subject to change at any time.

TRL provides judges to easily compare two completions.

Make sure to have installed the required dependencies by running:

pip install trl[judges]

Using the provided judges

TRL provides several judges out of the box. For example, you can use the HfPairwiseJudge to compare two completions using a pre-trained model from the Hugging Face model hub:

from trl import HfPairwiseJudge

judge = HfPairwiseJudge()
judge.judge(
    prompts=["What is the capital of France?", "What is the biggest planet in the solar system?"],
    completions=[["Paris", "Lyon"], ["Saturn", "Jupiter"]],
)  # Outputs: [0, 1]

Define your own judge

To define your own judge, we provide several base classes that you can subclass. For rank-based judges, you need to subclass BaseRankJudge and implement the BaseRankJudge.judge() method. For pairwise judges, you need to subclass BasePairJudge and implement the BasePairJudge.judge method. If you want to define a judge that doesn’t fit into these categories, you need to subclass BaseJudge and implement the BaseJudge.judge() method.

As an example, let’s define a pairwise judge that prefers shorter completions:

from trl import BasePairwiseJudge

class PrefersShorterJudge(BasePairwiseJudge):
    def judge(self, prompts, completions, shuffle_order=False):
        return [0 if len(completion[0]) > len(completion[1]) else 1 for completion in completions]

You can then use this judge as follows:

judge = PrefersShorterJudge()
judge.judge(
    prompts=["What is the capital of France?", "What is the biggest planet in the solar system?"],
    completions=[["Paris", "The capital of France is Paris."], ["Jupiter is the biggest planet in the solar system.", "Jupiter"]],
)  # Outputs: [0, 1]

Provided judges

PairRMJudge

class trl.PairRMJudge

< source >

( )

LLM judge based on the PairRM model from AllenAI.

This judge uses the PairRM model to rank pairs of completions for given prompts. It’s designed for pairwise comparison of language model outputs. The PairRM model is loaded using the llm-blender library and runs on the default Accelerator device.

Attributes:

blender (llm_blender.Blender): An instance of the Blender class from llm-blender.

Example:

>>> pairrm_judge = PairRMJudge()
>>> prompts = ["Translate 'hello' to French", "What's the capital of Japan?"]
>>> completions = [["Bonjour", "Salut"], ["Kyoto", "Tokyo"]]
>>> results = pairrm_judge.judge(prompts, completions)
>>> print(results)  # [0, 1] (indicating the first completion is preferred for the first prompt and the second)

This class requires the llm-blender library to be installed. Install it with: pip install llm-blender.

judge

< source >

( prompts: List completions: List shuffle_order: bool = True return_scores: bool = False temperature: float = 1.0 ) → Union[List[int, float]]

Parameters

prompts (List[str]) — List of prompts to judge.
completions (List[List[str]]) — List of completion pairs for each prompt.
shuffle_order (bool, optional, defaults to True) — Whether to shuffle the order of the completions to avoid positional bias.
return_scores (bool, optional, defaults to False) — If True, return probability scores of the first completion instead of ranks (i.e. a soft-judge).
temperature (float, optional, defaults to 1.0) — Temperature for scaling logits if return_scores is True.

Returns

Union[List[int, float]]

If return_scores is False, returns a list of ranks (0 or 1) for each prompt, indicating which completion is preferred. If return_scores is True, returns softmax probabilities for the first completion.

Raises

ValueError

ValueError — If the number of completions per prompt is not exactly 2.

Judge the completion pairs for the given prompts using the PairRM model.

Note: Unlike llm-blender, ranks are 0-indexed (0 means the first completion is preferred).

HfPairwiseJudge

class trl.HfPairwiseJudge

< source >

( model = 'meta-llama/Meta-Llama-3-70B-Instruct' token: Optional = None system_prompt: Optional = None )

Parameters

model (str, optional, defaults to "meta-llama/Meta-Llama-3-70B-Instruct") — Model to use for the judge.
token (str, optional) — Hugging Face API token to use for the huggingface_hub.InferenceClient.
system_prompt (str or None, optional, defaults to None) — The system prompt to be used for the judge. If not provided, a default prompt is used. Note that the system prompt should contain the following placeholders: {prompt}, {response0}, and {response1}. Also, the inference is called with max_tokens=1, consequently the system prompt should ask for a single token response.

Pairwise judge based on the Hugging Face API with chat completion.

This judge is relevant for assessing the quality chat models, where the completion is a response to a given prompt.

OpenAIPairwiseJudge

class trl.OpenAIPairwiseJudge

< source >

( model = 'gpt-4-turbo-preview' system_prompt: Optional = None max_requests: Optional = 1000 )

Parameters

model (str, optional, defaults to "gpt-4-turbo-preview") — Model to use for the judge.
system_prompt (str or None, optional, defaults to None) — System prompt to be used for the judge. If not provided, a default prompt is used. Note that the system prompt should contain the following placeholders: {prompt}, {response0}, and {response1}. Also, the inference is called with max_tokens=1, consequently the system prompt should ask for a single token response.
max_requests (int or None, optional, defaults to 1000) — Maximum number of requests to make to the OpenAI API. If set to None, there is no limit.

Judge based on the OpenAI API.

This judge is relevant for assessing the quality chat models, where the completion is a response to a given prompt.

AllTrueJudge

class trl.AllTrueJudge

< source >

( judges: List )

Parameters

judges (List[BaseBinaryJudge]) — A list of BaseBinaryJudge instances whose decisions will be unified.

Unify the decision of multiple BaseBinaryJudge instances.

Returns 1 only if all inner binary judges return 1. If any judge returns 0, it returns 0. If any judge returns -1, indicating a failure in its process, this judge will also return -1.

Implements the Mixture of Judges as described in the CGPO paper.

Base classes

BaseJudge

class trl.BaseJudge

< source >

( )

Base class for judges. The subclasses of this class should implement the judge method.

BaseBinaryJudge

class trl.BaseBinaryJudge

< source >

( )

Base class for binary judges.

judge

< source >

( prompts: List completions: List gold_completions: Optional = None shuffle_order: bool = True ) → List[int]

Parameters

prompts (List[str]) — List of prompts.
completions (List[str]) — List of completions.
gold_completions (List[str], optional) — List of gold completions if it exists.
shuffle_order (bool) — Whether to shuffle the order of the completions to avoid positional bias.

Returns

List[int]

A list of binary labels:

1 indicates that the completion satisfies the evaluated constraint.
0 indicates that the completion does not satisfy the evaluated constraint.

Judge the completion for a given prompt. Used to assess if a completion satisfies a constraint.

This base class should be used to implement binary evaluations as done in section 4.1.4 of the CGPO paper. It is relevant for assessing whether or not a prompt completion pair satisfies a specific contraint.

Note: If the judge returns -1 for any prompt, it indicates that the inner process used to compute the preference has failed. For instance, this could occur if the underlying language model or rule based contraint returned an invalid answer. In such cases, the caller should handle these invalid indices appropriately, possibly by implementing fallback logic or error handling.

BaseRankJudge

class trl.BaseRankJudge

< source >

( )

Base class for LLM ranking judges.

Example:

class MyRankJudge(BaseRankJudge):
    def judge(self, prompts, completions, shuffle_order=True):
        return ...  # Your ranking logic here

judge = MyRankJudge()
judge.judge(
    prompts=["The capital of France is", "The capital of Germany is"],
    completions=[[" Paris", " Marseille", "Lyon"], [" Munich", " Berlin"]]
)  # [[0, 1, 2], [1, 0]]

judge

< source >

( prompts: List completions: List shuffle_order: bool = True ) → List[List[int]]

Parameters

prompts (List[str]) — List of prompts.
completions (List[List[str]]) — List of completions list, where each element is a list of completions for the corresponding prompt.
shuffle_order (bool, optional, defaults to True) — Whether to shuffle the order of the completions to avoid positional bias.

Returns

List[List[int]]

List of lists of idxs, where each list contains the ranks of the completions for the corresponding prompt. E.g., [1, 2, 0] means that the second completion (idx=1) is the best, followed by the third, and then the first.

Judge the completion for the given prompts and return the ranks of each completion.

BasePairwiseJudge

class trl.BasePairwiseJudge

< source >

( )

Base class for pairwise judges.

judge

< source >

( prompts: List completions: List shuffle_order: bool = True ) → List[int]

Parameters

prompts (List[str]) — List of prompts.
completions (List[List[str]]) — List of completions pairs, where each element is a pair of completions for the corresponding prompt.
shuffle_order (bool, optional, defaults to True) — Whether to shuffle the order of the completions to avoid positional bias.

Returns

List[int]

List of idxs, where each idx is the rank of the best completion for the corresponding prompt. E.g., 1 means that the second completion (idx=1) is the best.

Judge the completion pairs for the given prompts.

Note: If the judge returns -1 for any prompt, it indicates that the inner process used to compute the preference has failed. For instance, this could occur if the underlying language model returned an invalid answer. In such cases, the caller should handle these invalid indices appropriately, possibly by implementing fallback logic or error handling.

< > Update on GitHub