Add timeh2o-v1-tiny wrapper and fev-bench submission

[Far0n]

Hi,

We'd like to submit TimeH2O-v1-tiny to fev-bench.
It's a 5.8M-parameter univariate time-series foundation model that runs through the h2ofm package.

On the full benchmark it gets a SQL skill score of 45.5, with a median runtime of 0.19 s per 100 series on a single RTX 5090.

This PR adds:

• models/timeh2o-v1-tiny/: the wrapper and requirements.txt (h2ofm==0.42.0)
• benchmarks/fev_bench/results/timeh2o-v1-tiny.csv: results for all 100 tasks
• an entry in models.yaml

To reproduce:

uv run python models/evaluate.py -m timeh2o-v1-tiny -b benchmarks/fev_bench/tasks.yaml

A few notes:

• The weights are gated on Hugging Face, so access has to be requested on the model page before running.
• h2ofm requires Python 3.12 or newer (we used 3.14).
• This checkpoint forecasts each target independently and does not use covariates.

The submission tests pass locally, and the results match our own internal fev-bench run.

Thanks for maintaining the benchmark!

Cheers,
Far0n

[shchur]

Great results, thanks for the contribution!

[Far0n]

Great results, thanks for the contribution!

Thank you!