datasys_trace.py

"""Hugging Face Datasets loader for the DataSys LLM serving trace."""

from __future__ import annotations

import json
from pathlib import Path

import datasets


_DESCRIPTION = """\
An anonymized LLM serving request trace for systems research on request
patterns, token accounting, latency, and token-bucket reuse.
"""

_HOMEPAGE = ""
_LICENSE = "other"


class DataSysTraceConfig(datasets.BuilderConfig):
    """Builder config for one JSONL file in the trace package."""

    def __init__(self, *, filename: str, features: datasets.Features, **kwargs):
        super().__init__(**kwargs)
        self.filename = filename
        self.features = features


_MODEL_PARAMETERS = {
    "temperature": datasets.Value("float64"),
    "max_tokens": datasets.Value("string"),
    "top_p": datasets.Value("float64"),
    "frequency_penalty": datasets.Value("float64"),
    "presence_penalty": datasets.Value("float64"),
    "seed": datasets.Value("int64"),
}

_TRACE_FEATURES = datasets.Features(
    {
        "id": datasets.Value("string"),
        "status": datasets.Value("string"),
        "created_at": datasets.Value("string"),
        "finished_at": datasets.Value("string"),
        "model": datasets.Value("string"),
        "model_parameters": _MODEL_PARAMETERS,
        "reported_token_input": datasets.Value("int64"),
        "reported_token_output": datasets.Value("int64"),
    }
)

_QWEN_BUCKET_FEATURES = datasets.Features(
    {
        **_TRACE_FEATURES,
        "token_count": datasets.Value("int64"),
        "bucket_ids": datasets.Sequence(datasets.Value("int64")),
    }
)


def _coerce_record(record: dict, include_buckets: bool) -> dict:
    """Fill optional nested keys so strict feature schemas stay stable."""
    model_parameters = record.get("model_parameters") or {}
    record["model_parameters"] = {
        "temperature": model_parameters.get("temperature"),
        "max_tokens": (
            None
            if model_parameters.get("max_tokens") is None
            else str(model_parameters.get("max_tokens"))
        ),
        "top_p": model_parameters.get("top_p"),
        "frequency_penalty": model_parameters.get("frequency_penalty"),
        "presence_penalty": model_parameters.get("presence_penalty"),
        "seed": model_parameters.get("seed"),
    }
    if include_buckets:
        record["bucket_ids"] = record.get("bucket_ids") or []
    return record


class DataSysTrace(datasets.GeneratorBasedBuilder):
    """DataSys trace dataset builder."""

    VERSION = datasets.Version("1.0.0")
    DEFAULT_CONFIG_NAME = "trace"

    BUILDER_CONFIGS = [
        DataSysTraceConfig(
            name="trace",
            version=VERSION,
            description="Full request-level trace metadata.",
            filename="trace.jsonl",
            features=_TRACE_FEATURES,
        ),
        DataSysTraceConfig(
            name="qwen3-32b-buckets",
            version=VERSION,
            description="Qwen/Qwen3-32B subset with token-bucket IDs.",
            filename="qwen3-32b-buckets.jsonl",
            features=_QWEN_BUCKET_FEATURES,
        ),
    ]

    def _info(self) -> datasets.DatasetInfo:
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=self.config.features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
        )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        data_file = dl_manager.download_and_extract(self.config.filename)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"filepath": data_file},
            )
        ]

    def _generate_examples(self, filepath: str):
        include_buckets = "bucket_ids" in self.config.features
        with Path(filepath).open("r", encoding="utf-8") as handle:
            for idx, line in enumerate(handle):
                line = line.strip()
                if not line:
                    continue
                yield idx, _coerce_record(json.loads(line), include_buckets)