Datasets:
Add tabSutra_39M_v1.0 (history only)
tabSutra_39M_v1.0 (history only)
This pull request adds the folder eval/tabSutra_39M_v1.0/ with tabSutra_39M_v1.0.parquet, which covers all 10,000
instances in the 78 dataset–term groups, and config.json, which declares the category "TSFM" and the input "History".
The method
The submission is history-only and zero-shot. tabSutra_39M_v1.0 is a 39M-parameter foundation model for structured
data that was pretrained on synthetic data only. Its model page is https://huggingface.co/tabsutra/tabSutra_39M_v1.0.
The inputs that were read
For each instance, the pipeline read the history, the n_horizon and the frequency columns. The
structured effects in macro and micro were not read, and neither was the tsfm_average_forecast column.
Timestamps
The benchmark has no dates, so each history was given synthetic timestamps at the benchmark's frequency, with row k
placed at k × step from an origin of zero, which is the Unix epoch.
Because the model derives phases from timestamps, those phases are positional rather than real dates.
The forecast
The forecast is the median of the model's 19 quantiles, with exactly n_horizon values per instance. Two
post-processing steps are applied, each decided per series from its own history. When the history has no negative
value, every quantile is floored at zero. When the history consists of whole numbers only, the median is rounded to the
nearest whole number.
MASE
The mase column follows the benchmark's definition. It is the mean absolute error over the horizon divided by the
mean absolute seasonal difference of the history, where the seasonal period comes fromgluonts.time_feature.get_seasonality and a one-step difference is used when the history is shorter than the period.
Recomputing the published moirai_2 results with this formula reproduces their mase column exactly on all 10,000
rows.
Training data
The training data was synthetic only. No GIFT-Eval, fev-bench, TIME or BOOM dataset was used in training in any form,
and no real-world time series was used as well.
Contact: https://tabsutra.com