Model’s Output

All models will generate an ouput per Doc supplied to the generation or loglikelihood fuctions.

class lighteval.models.model_output.ModelResponse

( input: str | list | None = None text: list = <factory> logprobs: list = <factory> argmax_logits_eq_gold: list = <factory> logits: list[list[float]] | None = None truncated_tokens_count: int = 0 padded_tokens_count: int = 0 input_tokens: list = <factory> output_tokens: list = <factory> unconditioned_logprobs: typing.Optional[list[float]] = None )

Parameters

input (str | list | None) — The original input prompt or context that was fed to the model. Used for debugging and analysis purposes.
text (list[str]) — The generated text responses from the model. Each element represents one generation (useful when num_samples > 1). Required for: Generative metrics, exact match, llm as a judge, etc.
logprobs (list[float]) — Log probabilities of the generated tokens or sequences. Required for: loglikelihood and perplexity metrics.
argmax_logits_eq_gold (list[bool]) — Whether the argmax logits match the gold/expected text. Used for accuracy calculations in multiple choice and classification tasks. Required for: certain loglikelihood metrics.
unconditioned_logprobs (Optional[list[float]]) — Log probabilities from an unconditioned model (e.g., without context). Used for PMI (Pointwise Mutual Information) normalization. Required for: PMI metrics.

A class to represent the response from a model during evaluation.

This dataclass contains all the information returned by a model during inference, including generated text, log probabilities, token information, and metadata. Different attributes are required for different types of evaluation metrics.

Usage Examples:

For generative tasks (text completion, summarization):

response = ModelResponse(
    text=["The capital of France is Paris."],
    input_tokens=[1, 2, 3, 4],
    output_tokens=[[5, 6, 7, 8]]
)

For multiple choice tasks:

response = ModelResponse(
    logprobs=[-0.5, -1.2, -2.1, -1.8],  # Logprobs for each choice
    argmax_logits_eq_gold=[False, False, False, False],  # Whether correct choice was selected
    input_tokens=[1, 2, 3, 4],
    output_tokens=[[5], [6], [7], [8]]
)

For perplexity calculation:

response = ModelResponse(
    text=["The model generated this text."],
    logprobs=[-1.2, -0.8, -1.5, -0.9, -1.1],  # Logprobs for each token
    input_tokens=[1, 2, 3, 4, 5],
    output_tokens=[[6], [7], [8], [9], [10]]
)

For PMI analysis:

response = ModelResponse(
    text=["The answer is 42."],
    logprobs=[-1.1, -0.9, -1.3, -0.7],  # Conditioned logprobs
    unconditioned_logprobs=[-2.1, -1.8, -2.3, -1.5],  # Unconditioned logprobs
    input_tokens=[1, 2, 3, 4],
    output_tokens=[[5], [6], [7], [8]]
)

Notes:

For most evaluation tasks, only a subset of attributes is required
The text attribute is the most commonly used for generative tasks
logprobs are essential for probability-based metrics like perplexity
argmax_logits_eq_gold is specifically for certain multiple choice/classification tasks
Token-level attributes (input_tokens, output_tokens) are useful for debugging
Truncation and padding counts help understand model behavior with long inputs

< > Update on GitHub

Lighteval

Model’s Output

class lighteval.models.model_output.ModelResponse