lvkaokao/benchmark_data

TaskWatch history

This dataset is written automatically by TaskWatch. It contains per-machine history for benchmark and shell jobs.

Layout

machines/<machine_id>/history.db      # raw SQLite source
machines/<machine_id>/manifest.json   # machine metadata and row counts
machines/<machine_id>/jobs.csv        # one row per job (latest state)
machines/<machine_id>/events.csv      # status-change events
machines/<machine_id>/snapshots.csv   # periodic progress/accuracy snapshots
machines/<machine_id>/logs/<job>/...  # optional uploaded per-run logs

The CSV files are generated for the Hugging Face dataset viewer; they are overwritten on every sync. The SQLite file remains the source of truth.

Tables

jobs.csv

machine, job_id, name, kind, status, phase, first_seen, last_seen, current_run, runs_total, active, event_count, snapshot_count, latest_accuracy, pid, command.

events.csv

machine, ts, job_id, name, kind, event, previous_status, status, phase, current_run, runs_total, pid, user, command, model, dataset.

Event types: discovered, status_change, reappeared, disappeared.

snapshots.csv

machine, ts, job_id, name, kind, status, phase, current_run, runs_total, accuracy, resolved, submitted, command.

Notes

• One machine owns machines/<machine_id>/; files are replaced on push.
• History may contain local paths, commands and hostnames. Use a private dataset repo if that is sensitive.
• The dataset card defines three viewer configs: jobs, events, snapshots. Each config combines all machines' CSV files; use the machine column to filter.
• If a config does not appear, open a CSV directly from the Files tab.