Skip to content

PrometheusEval

Critique and rank the quality of generations from an LLM using Prometheus 2.0.

PrometheusEval is a task created for Prometheus 2.0, covering both the absolute and relative evaluations. The absolute evaluation i.e. mode="absolute" is used to evaluate a single generation from an LLM for a given instruction. The relative evaluation i.e. mode="relative" is used to evaluate two generations from an LLM for a given instruction. Both evaluations provide the possibility of using a reference answer to compare with or withoug the reference attribute, and both are based on a score rubric that critiques the generation/s based on the following default aspects: helpfulness, harmlessness, honesty, factual-validity, and reasoning, that can be overridden via rubrics, and the selected rubric is set via the attribute rubric.

Note

The PrometheusEval task is better suited and intended to be used with any of the Prometheus 2.0 models released by Kaist AI, being: https://huggingface.co/prometheus-eval/prometheus-7b-v2.0, and https://huggingface.co/prometheus-eval/prometheus-8x7b-v2.0. The critique assessment formatting and quality is not guaranteed if using another model, even though some other models may be able to correctly follow the formatting and generate insightful critiques too.

Attributes

  • mode: the evaluation mode to use, either absolute or relative. It defines whether the task will evaluate one or two generations.

  • rubric: the score rubric to use within the prompt to run the critique based on different aspects. Can be any existing key in the rubrics attribute, which by default means that it can be: helpfulness, harmlessness, honesty, factual-validity, or reasoning. Those will only work if using the default rubrics, otherwise, the provided rubrics should be used.

  • rubrics: a dictionary containing the different rubrics to use for the critique, where the keys are the rubric names and the values are the rubric descriptions. The default rubrics are the following: helpfulness, harmlessness, honesty, factual-validity, and reasoning.

  • reference: a boolean flag to indicate whether a reference answer / completion will be provided, so that the model critique is based on the comparison with it. It implies that the column reference needs to be provided within the input data in addition to the rest of the inputs.

  • _template: a Jinja2 template used to format the input for the LLM.

Input & Output Columns

graph TD
    subgraph Dataset
        subgraph Columns
            ICOL0[instruction]
            ICOL1[generation]
            ICOL2[generations]
            ICOL3[reference]
        end
        subgraph New columns
            OCOL0[feedback]
            OCOL1[result]
            OCOL2[model_name]
        end
    end

    subgraph PrometheusEval
        StepInput[Input Columns: instruction, generation, generations, reference]
        StepOutput[Output Columns: feedback, result, model_name]
    end

    ICOL0 --> StepInput
    ICOL1 --> StepInput
    ICOL2 --> StepInput
    ICOL3 --> StepInput
    StepOutput --> OCOL0
    StepOutput --> OCOL1
    StepOutput --> OCOL2
    StepInput --> StepOutput

Inputs

  • instruction (str): The instruction to use as reference.

  • generation (str, optional): The generated text from the given instruction. This column is required if mode=absolute.

  • generations (List[str], optional): The generated texts from the given instruction. It should contain 2 generations only. This column is required if mode=relative.

  • reference (str, optional): The reference / golden answer for the instruction, to be used by the LLM for comparison against.

Outputs

  • feedback (str): The feedback explaining the result below, as critiqued by the LLM using the pre-defined score rubric, compared against reference if provided.

  • result (Union[int, Literal["A", "B"]]): If mode=absolute, then the result contains the score for the generation in a likert-scale from 1-5, otherwise, if mode=relative, then the result contains either "A" or "B", the "winning" one being the generation in the index 0 of generations if result='A' or the index 1 if result='B'.

  • model_name (str): The model name used to generate the feedback and result.

Examples

Critique and evaluate LLM generation quality using Prometheus 2_0

from distilabel.steps.tasks import PrometheusEval
from distilabel.models import vLLM

# Consider this as a placeholder for your actual LLM.
prometheus = PrometheusEval(
    llm=vLLM(
        model="prometheus-eval/prometheus-7b-v2.0",
        chat_template="[INST] {{ messages[0]"content" }}\n{{ messages[1]"content" }}[/INST]",
    ),
    mode="absolute",
    rubric="factual-validity"
)

prometheus.load()

result = next(
    prometheus.process(
        [
            {"instruction": "make something", "generation": "something done"},
        ]
    )
)
# result
# [
#     {
#         'instruction': 'make something',
#         'generation': 'something done',
#         'model_name': 'prometheus-eval/prometheus-7b-v2.0',
#         'feedback': 'the feedback',
#         'result': 6,
#     }
# ]

Critique for relative evaluation

from distilabel.steps.tasks import PrometheusEval
from distilabel.models import vLLM

# Consider this as a placeholder for your actual LLM.
prometheus = PrometheusEval(
    llm=vLLM(
        model="prometheus-eval/prometheus-7b-v2.0",
        chat_template="[INST] {{ messages[0]"content" }}\n{{ messages[1]"content" }}[/INST]",
    ),
    mode="relative",
    rubric="honesty"
)

prometheus.load()

result = next(
    prometheus.process(
        [
            {"instruction": "make something", "generations": ["something done", "other thing"]},
        ]
    )
)
# result
# [
#     {
#         'instruction': 'make something',
#         'generations': ['something done', 'other thing'],
#         'model_name': 'prometheus-eval/prometheus-7b-v2.0',
#         'feedback': 'the feedback',
#         'result': 'something done',
#     }
# ]

Critique with a custom rubric

from distilabel.steps.tasks import PrometheusEval
from distilabel.models import vLLM

# Consider this as a placeholder for your actual LLM.
prometheus = PrometheusEval(
    llm=vLLM(
        model="prometheus-eval/prometheus-7b-v2.0",
        chat_template="[INST] {{ messages[0]"content" }}\n{{ messages[1]"content" }}[/INST]",
    ),
    mode="absolute",
    rubric="custom",
    rubrics={
        "custom": "[A]\nScore 1: A\nScore 2: B\nScore 3: C\nScore 4: D\nScore 5: E"
    }
)

prometheus.load()

result = next(
    prometheus.process(
        [
            {"instruction": "make something", "generation": "something done"},
        ]
    )
)
# result
# [
#     {
#         'instruction': 'make something',
#         'generation': 'something done',
#         'model_name': 'prometheus-eval/prometheus-7b-v2.0',
#         'feedback': 'the feedback',
#         'result': 6,
#     }
# ]

Critique using a reference answer

from distilabel.steps.tasks import PrometheusEval
from distilabel.models import vLLM

# Consider this as a placeholder for your actual LLM.
prometheus = PrometheusEval(
    llm=vLLM(
        model="prometheus-eval/prometheus-7b-v2.0",
        chat_template="[INST] {{ messages[0]"content" }}\n{{ messages[1]"content" }}[/INST]",
    ),
    mode="absolute",
    rubric="helpfulness",
    reference=True,
)

prometheus.load()

result = next(
    prometheus.process(
        [
            {
                "instruction": "make something",
                "generation": "something done",
                "reference": "this is a reference answer",
            },
        ]
    )
)
# result
# [
#     {
#         'instruction': 'make something',
#         'generation': 'something done',
#         'reference': 'this is a reference answer',
#         'model_name': 'prometheus-eval/prometheus-7b-v2.0',
#         'feedback': 'the feedback',
#         'result': 6,
#     }
# ]

References