Add a proper README

README.md

@@ -1,4 +1,15 @@
 ---
+license: cc-by-4.0
+pretty_name: KernelSynth (annotated)
+task_categories:
+- time-series-forecasting
+size_categories:
+- 1M<n<10M
+tags:
+- synthetic
+- gaussian-process
+- kernel-synth
+- interpretability
 dataset_info:
   features:
   - name: start
@@ -21,3 +32,53 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# KernelSynth (annotated)
+
+One million synthetic univariate time series, each 1024 points long, drawn from a Gaussian
+process prior whose kernel is a random composition of up to five base kernels. This is the
+KernelSynth procedure from [Chronos](https://arxiv.org/abs/2403.07815) with one addition:
+**the generating kernel is kept alongside each series**. The ground-truth structure behind
+every series is therefore known, which makes the corpus usable for interpretability work
+rather than only for pretraining.
+
+## Fields
+
+- `start` — constant `2000-01-01T00:00:00`. The time index is arbitrary and carries no meaning.
+- `target` — the series itself: 1024 `float64` values sampled from the GP prior.
+- `selected_kernel_reprs` — the base kernels drawn from the kernel bank, as scikit-learn reprs.
+- `kernel_formula` — the composed kernel with explicit precedence, e.g.
+  `RBF(length_scale=1) * (DotProduct(sigma_0=1) + WhiteKernel(noise_level=1))`.
+  scikit-learn's own repr omits parentheses, so `(a + b) * c` and `a + b * c` are
+  indistinguishable there; this field disambiguates them.
+
+## Generation
+
+Produced by a script adapted from Chronos'
+[`kernel-synth.py`](https://github.com/amazon-science/chronos-forecasting/blob/main/scripts/kernel-synth.py),
+with these settings: 1,000,000 series, length 1024, at most 5 base kernels each, seed 1.
+Kernels are drawn with replacement from the 33-entry Chronos kernel bank and combined
+pairwise with random `+` / `*` operators. Each series draws from its own independent random
+stream derived from the seed, so the corpus is reproducible and independent of worker count.
+
+One deviation from upstream: a jitter of `1e-8 * mean(diag(cov))` is added to the covariance
+diagonal to keep it numerically positive semi-definite, since composed kernels are often
+ill-conditioned.
+
+## License and attribution
+
+The data is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).
+
+It was produced by a script adapted from Chronos' `kernel-synth.py` (Copyright Amazon.com,
+Inc., Apache-2.0). That license covers the generator code, not the series it emits, and no
+Chronos data is contained here — every series is sampled fresh from a GP prior. The method,
+however, is theirs; please cite the Chronos paper:
+
+```bibtex
+@article{ansari2024chronos,
+  author  = {Ansari, Abdul Fatir and Stella, Lorenzo and Turkmen, Caner and Zhang, Xiyuan and Mercado, Pedro and Shen, Huibin and Shchur, Oleksandr and Rangapuram, Syama Syndar and Pineda Arango, Sebastian and Kapoor, Shubham and Zschiegner, Jasper and Maddix, Danielle C. and Wang, Hao and Mahoney, Michael W. and Torkkola, Kari and Gordon Wilson, Andrew and Bohlke-Schneider, Michael and Wang, Yuyang},
+  title   = {Chronos: Learning the Language of Time Series},
+  journal = {arXiv preprint arXiv:2403.07815},
+  year    = {2024}
+}
+```