Vita-Mojo/kitchensim-v1

KitchenSim — Synthetic Restaurant Operations Dataset (v1)

A parametric simulator-generated event stream for 60 simulated restaurant kitchens across 7 days. Designed for benchmarking demand forecasting, operational-state inference, and self-supervised representation learning on time-series event data.

This is the publish-ready version. A hidden ground-truth state column (hidden_true_state) was used during the original simulator's classifier and is stripped here — only what a downstream consumer would see in real operational data is included.

Schema

Statistics

• —1,163,151 events
• —60 stores × 7 days (Mon 2024-01-01 → Sun 2024-01-07 UTC)
• —~16 MB compressed (zstd) parquet
• —193,837 orders generated
• —Reproducibility: byte-identical given the master seed in the code repository generator config

Store types

Five archetypal kitchens with distinct operational profiles:

Dynamics

• —Non-homogeneous Poisson arrivals modulated by hour-of-day (double-peak at lunch + dinner) and day-of-week (Friday peak, Sunday dip).
• —State machine per store: stable / degraded / collapsed, classified by queue_depth / base_capacity and recent_fulfilment_EMA / median_fulfilment.
• —Non-linear load_factor 1 + 0.6 * max(0, q/cap - 1)^1.5 (capped at 5.0) drives the kitchen-spiral dynamic that's the signal of interest.
• —Stochastic disruptions per store per day: equipment failure (3%, capacity -40% for 8-15 min), demand spike (5%, arrival rate ×2 for 5-10 min), staff shortage (2%, capacity -25% for 30-90 min).

Loading

import pandas as pd
df = pd.read_parquet("hf://datasets/Vita-Mojo/kitchensim-v1/synthetic_ops_v1_publish.parquet")
print(df.head())

Use cases

• —Self-supervised pretraining of sequence encoders (this dataset trained Vita-Mojo/opswm1)
• —Benchmarking demand-forecasting models with realistic daily/weekly seasonality
• —Operational state-inference research (the simulator's classifier rule is documented in the model card so you can compare your inferred states to ground truth on the training split if you re-generate with the seed)
• —Evaluation of representation learning for tabular event streams

Limitations

• —Synthetic: dynamics are mathematically generated, not observed. The store-type taxonomy, arrival curves, and disruption mix are stylised abstractions of real-world patterns, not measured ones.
• —No multi-week seasonality: only 7 days, so weekly cycles are observed once. The day-of-week multiplier captures relative patterns; absolute calibration to a real chain is undefined.
• —No price/menu dynamics: prices are a simple £3 + £4 × items_total + N(0, £1.50) formula. Real-world discounting and menu mix are absent.
• —No customer-side state: the dataset only contains kitchen-side events. Customer wait times, abandonments, complaints are not modelled.

License

Apache 2.0