---
license: apache-2.0
pretty_name: skycmd synthetic time series (s6-cast generators)
task_categories:
- time-series-forecasting
tags:
- skycmd
configs:
- config_name: gp
  data_files: data/gp.parquet
- config_name: mixed
  data_files: data/mixed.parquet
- config_name: pulse
  data_files: data/pulse.parquet
- config_name: sarima
  data_files: data/sarima.parquet
- config_name: tsi
  data_files: data/tsi.parquet
---

# skycmd synthetic time series (s6-cast generators)

A fixed sample of the four synthetic series generators that pretrain the s6-cast forecaster (Gaussian processes, pulses, trend-seasonality-irregular, SARIMA), and of their training mix. Generated by skycmd's own Rust code: no third-party data.

Part of [skycmd](https://github.com/maxgfr/skycmd), tiny decision models trained from scratch in Rust that run in the
browser. Written by `sky-train cast-synth --kind <family> --n 1000 --len 2048 --seed 1, and --n 4000 for the mix` at commit [`8cc8e276`](https://github.com/maxgfr/skycmd/commit/8cc8e276adabbe1cc37998fb70fdd5c947ffda81).
Licences of every set: [DATA_LICENSES.md](https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md).

Each row is one synthetic series: `id` (`<family>-<i>`) and `values` (2,048 floats). The generators
are ours (`crates/sky-infer/src/cast/synth.rs`), after Reverso (arXiv 2602.17634, appendix B) and
Simulation Alone (arXiv 2601.00970), and deterministic per seed. s6-cast does not train on these
rows: it trains on an endless stream from the same generators (a refreshed pool, other seeds).
This repository is a fixed sample of that stream (`sky-train cast-synth --seed 1 --len 2048`), to
inspect the generators or to pretrain another forecaster on the same distribution.

## Sources and licences

Each config is one source and keeps **its source's licence** (share-alike sources stay share-alike).
Attribution goes to the upstream authors; the conversion is ours.

| config | upstream | licence | how it is used | rows |
|---|---|---|---|---:|
| `gp` | ours: `sky_infer::cast::synth`, family `gp` | apache-2.0 | KernelSynth-style Gaussian-process draws: 1 to 5 kernels (constant, linear, RBF, rational quadratic, Matérn, periodic) combined by sum or product, sampled with random Fourier features | 1,000 |
| `mixed` | ours: the training mix | apache-2.0 | the four families drawn with weights 0.3 / 0.15 / 0.3 / 0.25, as in s6-cast training | 4,000 |
| `pulse` | ours: family `pulse` | apache-2.0 | trapezoid pulse trains (spikes or dips) over a baseline, with noise | 1,000 |
| `sarima` | ours: family `sarima` | apache-2.0 | seasonal ARIMA with stationary AR parts (sampled through their partial autocorrelations), optional differencing | 1,000 |
| `tsi` | ours: family `tsi` | apache-2.0 | trend (linear, exponential, polynomial, piecewise) + up to 3 seasonal waves + Gaussian, Laplace or Student-t noise + outliers + level shifts | 1,000 |

## Files

| file | rows |
|---|---:|
| `data/gp.parquet` | 1,000 |
| `data/mixed.parquet` | 4,000 |
| `data/pulse.parquet` | 1,000 |
| `data/sarima.parquet` | 1,000 |
| `data/tsi.parquet` | 1,000 |

```python
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-cast-synth", "gp")
```

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream
sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.