Datasets:

eth-easl
/

swissai-serving-trace

pretty_name: DataSys LLM Serving Trace
license: other
language:
  - en
tags:
  - llm-serving
  - inference
  - request-trace
  - token-reuse
  - qwen3
size_categories:
  - 10M<n<100M

LLM Serving Trace

This dataset contains anonymized LLM serving request metadata prepared for systems research on request patterns, token accounting, latency, and reusable input-token buckets. It is packaged as newline-delimited JSON.

Dataset Files

File	Rows	Size	Description
`trace.jsonl`	16,329,237	6.98 GB	Request-level trace records with timestamps, status, model name, reported token counts, and generation parameters.
`qwen3-32b-buckets.jsonl`	3,994,435	4.62 GB	Qwen/Qwen3-32B subset augmented with token-bucket information for token reuse analysis.

Loading

The files are distributed as raw newline-delimited JSON. The model_parameters field is heterogeneous across requests (mixed types and occasional extra keys), so the schema is not friendly to datasets.load_dataset's automatic Arrow inference. Download the files with huggingface_hub and read them directly.

import json
from pathlib import Path

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="eth-easl/swissai-serving-trace",
    repo_type="dataset",
    allow_patterns=["trace.jsonl"],  # or ["qwen3-32b-buckets.jsonl"], or both
)

with (Path(local_dir) / "trace.jsonl").open("r", encoding="utf-8") as handle:
    for line in handle:
        record = json.loads(line)
        # process record

To grab a single file without materialising the whole snapshot tree, use hf_hub_download:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="eth-easl/swissai-serving-trace",
    filename="qwen3-32b-buckets.jsonl",
    repo_type="dataset",
)

If you prefer a dataframe and have enough memory, pandas can read the JSONL directly (trace.jsonl is ~7 GB, so consider chunksize=):

import pandas as pd

frames = pd.read_json(path, lines=True, chunksize=100_000)
for chunk in frames:
    ...  # process chunk

Schema

`trace`

Each row contains:

Field	Type	Description
`id`	string	Request identifier.
`status`	string	Request status, for example `DEFAULT` or `ERROR`.
`created_at`	string	Request creation timestamp in ISO-8601 format.
`finished_at`	string	Request completion timestamp in ISO-8601 format.
`model`	string	Model identifier used for the request.
`model_parameters`	object	Generation parameters supplied with the request.
`reported_token_input`	int64	Reported input-token count, or `-1` when unavailable.
`reported_token_output`	int64	Reported output-token count, or `-1` when unavailable.

model_parameters is the raw user-supplied generation config and is deliberately not normalised. Most rows contain the keys below, but values are mixed JSON types because they reflect what the client sent:

Field	Observed JSON types
`temperature`	number, `"null"` string, missing
`max_tokens`	number, string, `null`
`top_p`	number, `"null"` string, missing
`frequency_penalty`	number, string, list, `null`, missing
`presence_penalty`	number, string, bool, `null`, missing
`seed`	integer, missing

A small number of rows also carry an extra n integer key. Coerce these to your preferred schema at read time.

`qwen3-32b-buckets`

This config contains the same request metadata fields as trace, plus:

Field	Type	Description
`token_count`	int64	Number of Qwen/Qwen3-32B input tokens retained for the request.
`bucket_ids`	list[int64]	Deterministic token-bucket identifiers for the request input. Buckets were generated with 16-token buckets and right padding.

Intended Uses

This dataset is intended for research on LLM serving workloads, including:

request arrival and completion patterns;
latency and status analysis;
input/output token accounting;
cache locality and token reuse over bucketized input-token sequences.

The bucketized Qwen subset is useful when the raw token IDs should not be distributed or when experiments only need stable identifiers for repeated token chunks.

Limitations

The trace records contain serving metadata rather than original prompts or model outputs. Reported token counts may be unavailable for failed or incomplete requests, represented as -1. The bucket_ids are model- and preprocessing- specific and should not be interpreted as vocabulary token IDs.