xpark.dataset.TextPredicateEval

class xpark.dataset.TextPredicateEval(predicate: str, /, *, base_url: str, model: str, api_key: str = 'NOT_SET', max_qps: int | None = None, max_concurrency: int | None = None, max_retries: int = 0, fallback_response: bool | None = None, cascade: CascadeConfig | None = None, hint: str | list[str] | None = None, **kwargs: Any)

TextPredicateEval processor evaluates whether input texts satisfy a given predicate condition.

This processor uses a Large Language Model (LLM) to determine if each text in a column matches the specified predicate, returning True or False for each input.

Parameters:

• predicate – The predicate to evaluate.

• base_url – The base URL of the LLM server.

• model – The request model name.

• api_key – The request API key.

• max_qps – The maximum query-per-second rate for remote LLM requests.

• max_concurrency – The maximum number of in-flight remote LLM requests allowed concurrently.

• max_retries – The maximum number of retries per request in the event of failures. We retry with exponential backoff upto this specific maximum retries.

• fallback_response – The response value to return when the LLM request fails. If set to None, the exception will be raised instead.

• cascade – Optional CascadeConfig for cascade mode. See CascadeConfig for details.

• hint – Optional extra instructions or constraints to guide the model (e.g. domain rules that refine how the predicate should be judged). Accepts either a single string or a list of strings, where each item is one hint written in plain text. Passing a list is recommended — use one string per hint. Do not include output-format rules in the hint, as they are injected automatically.

• **kwargs –

  Keyword arguments to pass to the openai.AsyncClient.chat.completions.create API. logprobs: If True, return a pa.StructArray containing both the

  prediction and per-token logprobs instead of a plain prediction string. Maps directly to the OpenAI logprobs parameter.

  top_logprobs: Number of most likely tokens to return at each position,

  maps directly to the OpenAI top_logprobs parameter. Only meaningful when logprobs is True.

Examples

from xpark.dataset.expressions import col
from xpark.dataset import TextPredicateEval, from_items

ds = from_items(["The iconic tower in the capital of France is illuminated with lights."])
ds = ds.with_column(
    "eval",
    TextPredicateEval(
        predicate="The text describes Paris",
        model="deepseek-v3-0324",
        base_url=os.getenv("LLM_ENDPOINT"),
        api_key=os.getenv("LLM_API_KEY"),
    )
    .options(num_workers={"IO": 1}, batch_size=1)
    .with_column(col("item")),
)

print(ds.take_all())

# Cascade mode: proxy model first, then forward uncertain samples to base model
import math
from xpark.dataset.utils import CascadeConfig, elementwise_cascade

@elementwise_cascade
def cascade_fn(text: str, logprobs: list[dict] | None) -> bool:
    if not logprobs:
        return True
    prob = math.exp(logprobs[0]["logprob"]) * 100
    return prob < 95.0  # Forward if confidence < 95%

ds = ds.with_column(
    "eval",
    TextPredicateEval(
        predicate="The text describes Paris",
        model="deepseek-v3-0324",
        base_url=os.getenv("LLM_ENDPOINT"),
        api_key=os.getenv("LLM_API_KEY"),
        cascade=CascadeConfig(
            proxy_model="Qwen2.5-3B-Instruct",
            proxy_base_url="http://local-vllm:8000/v1",
            cascade_factory=lambda: cascade_fn,
        ),
    )
    .options(num_workers={"IO": 1}, batch_size=1)
    .with_column(col("item")),
)

Methods

__call__(texts)	Call self as a function.
options(**kwargs)
with_column(texts)

__call__(texts: pa.ChunkedArray) → pa.Array | pa.StructArray

Call self as a function.

options(**kwargs: Unpack[ExprUDFOptions]) → Self

with_column(texts: pa.ChunkedArray) → pa.Array | pa.StructArray