File size: 2,685 Bytes
4dd7cf7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 | ---
tags:
- multabench
- embeddings
license: other
license_name: case-license
license_link: https://huggingface.co/SAP/CASE/blob/main/LICENSE
---
# MulTaBench row embeddings (CASE)
Part of [MulTaBench](https://github.com/alanarazi7/MulTaBench), a benchmark for multimodal tabular learning.
[CASE](https://github.com/SAP-samples/case) embeddings of the MulTaBench text datasets, from
[SAP/CASE](https://huggingface.co/SAP/CASE) (Gemma-3-12B) with `CaseTransformer`'s defaults, cached so the CASE models'
leaderboard runs don't re-encode them. CASE fills its KV cache with up to 128 of a fold's training rows and their
targets, so its embeddings are per fold: each `<dataset>/10k/fold_<k>/embeddings.npz` maps every distinct train and test
row of the fold, keyed by its cells joined by tabs, to its 3840-d embedding (float16; PCA is fitted per run, on the
training rows); `meta.json` holds the fold's rows and columns, the encoding time and the hardware. See
[Cached embeddings](https://github.com/alanarazi7/MulTaBench#cached-embeddings).
Folds whose context didn't fit CASE's batch of 16 on an 80 GB H100 were embedded in smaller batches, 9 of them without
cuDNN's attention (`git` 19aaab5 in their `meta.json`); this changes embeddings only at bf16 rounding (mean cosine 0.9999
to the default's, on 500 rows of two folds).
These are outputs of SAP/CASE, a Gemma 3 derivative licensed for research purposes only; see its
[licence](https://huggingface.co/SAP/CASE/blob/main/LICENSE), the
[Gemma Terms of Use](https://ai.google.dev/gemma/terms) and the
[Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
**Paper:** [MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image](https://huggingface.co/papers/2605.10616) · **Leaderboard:** [multabench/leaderboard](https://huggingface.co/spaces/multabench/leaderboard)
## Citation
```bibtex
@misc{arazi2026multabenchbenchmarkingmultimodaltabular,
title={MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image},
author={Alan Arazi and Eilam Shapira and Shoham Grunblat and Mor Ventura and Elad Hoffer and Gioia Blayer and David Holzmüller and Lennart Purucker and Gaël Varoquaux and Frank Hutter and Roi Reichart},
year={2026},
eprint={2605.10616},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2605.10616},
}
```
```bibtex
@inproceedings{case,
title={Enhancing Tabular Learners with Context-Aware Semantic Embeddings},
author={Günther Schindler and Maximilian Schambach and Johannes Höhne},
booktitle={Advances in Neural Information Processing Systems},
year={2026},
url={https://arxiv.org/abs/2608.03565},
}
```
|