## tabSutra_39M_v1.0 (history only)
Adds `eval/tabSutra_39M_v1.0/` with `tabSutra_39M_v1.0.parquet` (10,000 instances, 78 dataset–term groups) and `config.json` (category TSFM, input History).
**Method.** History-only, zero-shot. tabSutra_39M_v1.0 is a 39M-parameter foundation model for structured data, pretrained on synthetic data only. Model page: https://huggingface.co/tabsutra/tabSutra_39M_v1.0. The same checkpoint and the same post-processing are used in our GIFT-Eval and fev-bench submissions.
**Inputs read.** Each instance's `history`, `n_horizon` and `frequency`. The structured effects (`macro`, `micro`) and the `tsfm_average_forecast` column were not read.
**Timestamps.** The benchmark has no dates, so each history was given timestamps at the benchmark's frequency, with row k at k × step from an origin of zero (the Unix epoch). The origin is arbitrary and carries no information; only the step matters. The model derives phases from timestamps, so these phases are positional…
