multabench/embeddings-tile-legacy

MulTaBench text embeddings (TILE)

Part of MulTaBench, a benchmark for multimodal tabular learning.

Frozen TILE embeddings of the MulTaBench datasets (Llama-3.1-8B, first 14 layers), cached so leaderboard runs don't re-encode them. TILE reads whole rows, so each <dataset>/embeddings.npz maps every distinct serialized row to one mean-pooled 4096-d vector per feature (float16; PCA is fitted per run, on the training rows); meta.json holds the dataset's rows and columns, the encoding time per feature and the hardware. See Cached embeddings.

Paper: MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image · Leaderboard: multabench/leaderboard

Citation

@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},
}