Datasets:

harpertoken
/

stat

title: harpertoken/stat
library_name: pandas
tags:
  - system-monitoring
  - time-series
  - anomaly-detection
  - predictive-maintenance
  - macOS
license: mit
language:
  - en

harpertoken/stat

Summary

A single-session capture of live macOS system telemetry: 6,190 samples of CPU, memory, disk, network, and battery readings taken at one-second intervals over roughly 1 hour 44 minutes on 2025-03-02.

This is one machine, one continuous session, one workload. It is small enough to load in a second and structured for time-series experiments, not a benchmark.

Field	Type	Range	Notes
`timestamp`	timestamp	2025-03-02 10:24:07 → 12:07:52	1 s median interval, 2 s max
`cpu_usage`	list[8] of float	0.0 → 100.0	One value per logical core, 8 cores
`memory_used_mb`	float	2762.0 → 3585.7	Used RAM, not total
`disk_read_mb`	float	4,141,332 → 4,306,966	Cumulative counter, unit unverified
`disk_write_mb`	float	1,462,612 → 1,611,402	Cumulative counter, unit unverified
`net_sent_mb`	float	5,563.3 → 6,063.3	Cumulative counter, unit unverified
`net_recv_mb`	float	2,379.4 → 2,937.2	Cumulative counter, unit unverified
`battery_status`	float	15.0 → 35.0	Battery percentage; falls to 15 then recovers to 35
`cpu_temp`	null	none	Always null; see below

Three things are worth knowing before you use this.

The disk and network columns are since-boot counters, not per-interval activity. They only ever increase, and they start well above zero because the machine was already running when capture began. Across the session they advance by 165,634 (disk_read_mb), 148,790 (disk_write_mb), 500 (net_sent_mb) and 558 (net_recv_mb).

Treat those advances as raw counts, not confirmed volumes. The _mb suffix is unverifiable from the data alone: read as MB they imply ~95 GB/h of reads and ~85 GB/h of writes, which is high but possible; read as KB they imply ~95 MB/h, which also fits. Only a byte reading (162 KB total) is inconsistent with this machine's CPU profile. Confirm the unit with whatever produced the capture before reporting throughput in any physical unit.

Feed the raw values to a model and the dominant signal is "how long has this host been up," which correlates with the target and will inflate your results. Difference them first:

df["disk_read_delta"] = df["disk_read_mb"].diff()

cpu_temp is empty. All 6,190 rows are N/A; the temperature probe was not available on this machine. The column is retained so the schema matches the collector, but it carries no information and should be dropped rather than imputed.

Loading

load_dataset("harpertoken/stat") works and returns all 6,190 rows in a single train split, with cpu_usage already decoded into a list of eight floats and cpu_temp as None. The full CSV and a Parquet rebuilt from it are published alongside it; read one explicitly when you want to control the parsing:

import pandas as pd

df = pd.read_csv(
    "https://huggingface.co/datasets/harpertoken/stat/resolve/main/system_monitoring_dataset.csv"
)
df["timestamp"] = pd.to_datetime(df["timestamp"])
df["cpu_usage"] = df["cpu_usage"].apply(
    lambda v: [float(x) for x in v.strip("[]").split(",")]
)
df = df.drop(columns=["cpu_temp"])

# per-core CPU -> mean, and cumulative counters -> per-interval deltas
df["cpu_mean"] = df["cpu_usage"].apply(sum) / df["cpu_usage"].apply(len)
for col in ["disk_read_mb", "disk_write_mb", "net_sent_mb", "net_recv_mb"]:
    df[col.replace("_mb", "_delta")] = df[col].diff()

Prefer a parquet round-trip, which keeps the list column typed and skips the string parsing above. pyarrow.read_table will not open an HTTPS URL, so fetch the file first:

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download(
    "harpertoken/stat", "system_monitoring_dataset.parquet", repo_type="dataset"
)
df = pq.read_table(path).to_pandas()

# cpu_usage arrives as an array of 8 floats, and timestamp is already datetime

Intended uses

Short-horizon forecasting of memory and CPU on a single host, and unsupervised anomaly detection where an alert is raised on a burst of load or memory growth. Because the session is one hour on one machine, treat it as a demonstration of the collection format rather than evidence of generalization; a model tuned on this capture will need refitting on data from the deployment host.

Out of scope

Multimachine or fleet monitoring, multi-day trends, thermal analysis (cpu_temp is null), and battery-cycle modeling: the session spans a single discharge from 33% to 15% followed by a recharge to 35%, which is one partial charge cycle, not enough to model cycling behaviour.

Provenance

Collected locally on a Mac via a system-metric sampler. No network payloads, process names, or user data are present; the telemetry is host-level only. Licensed MIT.

Contact: coccinella.labs@icloud.com